Privacy Communication Patterns for Domestic Robots
家用机器人的隐私通信模式

Smart assistants have long become integral parts of many homes, as they make life more enjoyable by providing entertainment or supporting with daily chores. Most of these devices are either placed in a dedicated area, such as smart speakers or have minimal interaction capabilities, such as robot vacuums. Despite their restricted movement and interaction, they already cause various privacy concerns as their sensors collect and process sensitive data. Such concerns include the smart assistant transmitting data without explicit consent [26] or being exposed to microphones that are always listening and sharing recordings with third parties [27]. However, through advancements in AI and robotics,
智能助手早已成为许多家庭不可或缺的一部分，因为它们通过提供娱乐或支持日常家务劳动，让生活变得更加惬意。这些设备中的大多数要么被放置在专用区域（如智能扬声器），要么具有最低限度的交互能力（如机器人吸尘器）。尽管这些设备的移动和交互受到限制，但由于其传感器会收集和处理敏感数据，因此已经引发了各种隐私问题 。这些问题包括智能助手在未经明确同意的情况下传输数据[26]，或暴露在一直在监听并与第三方共享录音的麦克风中[27]。然而，随着人工智能和机器人技术的进步，这些问题也将得到解决、

future smart assistants will not remain static and passive (c.f., Amazon Astro). Quite the contrary - they will gain various locomotion and interaction capabilities, allowing them to enter all areas and even physically manipulate the environment. Such domestic robots will increase our convenience as they take over tasks like folding laundry or cleaning bathrooms. However, this will make them even more intrusive as the robots can access all rooms or even search through personal belongings, paving the way for various privacy concerns.
未来的智能助理将不再是静态和被动的（例如亚马逊 Astro）。恰恰相反，它们将获得各种运动和交互能力，能够进入所有区域，甚至对环境进行物理操控。这种家用机器人将为我们带来更多便利，因为它们将接手叠衣服或打扫浴室等工作。然而，这也将使它们更具侵扰性，因为机器人可以进入所有房间，甚至搜查个人物品，从而为各种隐私问题铺平道路。

Due to their advanced locomotion and interaction capabilities and potential for social bonding, domestic robots pose completely new threats to users' psychological, social, and physical privacy [32]. Users, for example, report being concerned about getting accidentally recorded while the robot moves past or interacts with other entities [28]. Moreover, humanoid robots pose a particular threat to users' privacy, as they provoke trust, leading to users' willingly sharing feelings and sensitive information [48]. Further, their humanoid appearance lets people underestimate their capabilities as they relate them to human capabilities [28]. As a result, experts demand that robots regularly communicate their privacy states to users, such as unambiguously indicating whether they are currently recording [24]. Even though there have been suggestions for such communication patterns [32], research is scarce and lacks an encompassing picture. Thus, we do not know which patterns evoke trust, are understandable, have good notification qualities, and are favored by users.
由于家用机器人具有先进的运动和交互能力以及潜在的社会联系，它们对用户的心理、社会和身体隐私构成了全新的威胁[32]。例如，用户担心机器人经过或与其他实体互动时被意外记录下来 [28]。此外，仿人机器人对用户的隐私构成了特别的威胁，因为它们会激起信任，导致用户愿意分享情感和敏感信息[48]。此外，仿人机器人的外观会让人低估它们的能力，因为人们会把它们与人类的能力联系起来 [28]。因此，专家们要求机器人定期向用户传达自己的隐私状态，比如明确表示自己是否正在录音 [24]。尽管已经有人提出了此类交流模式的建议 [32]，但相关研究并不多，而且缺乏全面的描述。因此，我们不知道哪些模式能唤起信任、易于理解、具有良好的通知品质，并且受到用户的青睐。

To close this gap, we first investigated the impact of locomotion and interaction capabilities on privacy concerns. Then, we investigated how domestic robots can communicate their sensor states to allow users to assess potential privacy risks. We explore two dimensions that contribute to privacy risks: a) the locomotion (4 levels) and b) interaction (3 levels) capabilities. We conducted an online survey to understand how the resulting scenarios affect user privacy concerns and investigated reasons for concerns. We then elicited communication patterns in three focus groups that allow users to assess the robot's sensor states (cameras, microphones, and network connectivity). Finally, we conducted
为了缩小这一差距，我们首先研究了运动和交互能力对隐私问题的影响。然后，我们研究了家用机器人如何传达其传感器状态，以便用户评估潜在的隐私风险。我们探讨了导致隐私风险的两个方面：a) 运动能力（4 级）和 b) 交互能力（3 级）。我们进行了一项在线调查 ，以了解由此产生的 场景如何影响用户的隐私顾虑，并调查顾虑的原因。然后，我们在三个焦点小组 中收集了交流模式，让用户评估机器人的传感器状态（摄像头、麦克风和网络连接）。最后，我们进行了
a large-scale survey ( ) to understand which patterns performed best regarding trust, privacy, understandability, notification qualities, and general user preference.
大规模调查（ ），以了解哪些模式在信任度、隐私性、易理解性、通知质量和一般用户偏好方面表现最佳。

This paper provides a path to allow domestic robots to enter our homes while keeping privacy concerns low. First, we found that advanced locomotion and interaction capabilities increase users' concerns. Second, we provide a set of 86 communication patterns to indicate the robots' microphone, camera, and connectivity states. Finally, we found that most of our elicited communication patterns scored equally well, showing that which pattern to use depends on the characteristics of the situation. To the best of our knowledge, this paper is the first to provide (1) an understanding of how increased locomotion and interaction capabilities of future smart assistants affect users' privacy concerns, (2) construct an encompassing set of various communication patterns for domestic robots to indicate the state of their privacy-relevant capabilities, and (3) provide insights into the quality of the communication patterns. Furthermore, we developed an interactive web application to facilitate the exploration, filtering, and retrieval of appropriate communication patterns based on designers' and researchers' diverse needs and preferences. With this, our set of patterns will guide developers and researchers in ensuring a privacy-preserving future with domestic robots.
本文提供了一条既能让家用机器人进入我们的家庭，又能降低隐私担忧的途径。首先，我们发现先进的运动和交互能力会增加用户的担忧。其次，我们提供了一组 86 种通信模式，用于指示机器人的麦克风、摄像头和连接状态。最后，我们发现大多数诱导出的交流模式得分都一样高，这表明使用哪种模式取决于情况的特点。据我们所知，本文首次提供了：（1）对未来智能助手运动和交互能力的增强如何影响用户隐私关注的理解；（2）为家用机器人构建了一套包罗万象的各种交流模式，以显示其隐私相关能力的状态；以及（3）对交流模式的质量提供了见解。此外，我们还开发了一个交互式网络应用程序，以便根据设计者和研究人员的不同需求和偏好，探索、筛选和检索适当的交流模式。有了这套模式，我们将指导开发人员和研究人员确保家用机器人在未来能够保护隐私。

First, we report on privacy in smart home contexts: The specific risks, users' concerns, and mitigation strategies. Second we highlight work on privacy concerns of domestic robots.
首先，我们报告了智能家居环境中的隐私问题：具体风险、用户关注的问题以及缓解策略。其次，我们将重点介绍家用机器人的隐私问题。

Through their placement in our intimate spaces, smart home devices are exceptionally prone to revealing sensitive information when exploited. Research, for example, showed how data from smart devices allows retracing identities [42], tracking user behavior [4], revealing the number of people in a household, or their sleeping and eating routines [40].
智能家居设备放置在我们的私密空间中，一旦被利用，极易泄露敏感信息。例如，研究表明，智能设备提供的数据可以追溯身份[42]、追踪用户行为[4]、揭示家庭中的人数或他们的睡眠和饮食规律[40]。

While some users are unable to pinpoint the concrete dangers posed by smart devices [20,34,35], they still feel a sense of unease or have concrete privacy concerns when in their vicinity [50]. Such concerns include personal data being revealed without explicit consent [26], for example, through always-listening smart speakers that share these data with third parties [27]. Prior research also found a diverging danger perception regarding different sensor types [50]. Users are most concerned about cameras and microphones [12, 50] and mostly consider temperature or motion sensors [50] less concerning. Some even express clear skepticism that these sensors cause any concern at all .
虽然有些用户无法明确指出智能设备带来的具体危险[20,34,35]，但他们在使用智能设备时仍会感到不安或有具体的隐私顾虑[50]。这些担忧包括个人数据在未经明确同意的情况下被泄露[26]，例如，通过始终监听的智能扬声器与第三方共享这些数据[27]。先前的研究还发现，不同类型的传感器存在不同的危险感知[50]。用户最担心的是摄像头和麦克风[12, 50]，而大多认为温度或运动传感器[50]不太危险。有些用户甚至明确表示怀疑这些传感器是否会引起任何问题 。

Prior research also investigated approaches to counter these concerns, including technological measures, such as implementing traffic shaping techniques [5], auto-configuring smart devices and implementing automatic updates [30], or introducing frameworks that automatically adjust the privacy level in smart homes depending on contexts [41] or pre-defined privacy zones [7]. Moreover, through co-design studies, Yao et al. [55] suggest different control mechanisms, such as disconnecting devices from the internet and keeping data local, increasing transparency and control, and providing access control through different modes. Next to these approaches, a more recent thread of research focuses on tangible control mechanisms [3, 14, 38, 52]. A major advantage of these mechanisms is their high understandability, which instills trust and guarantees inclusivity, especially for people with low technological understanding [3, 52]. Moreover, Chalhoub et al. [12] found that physical camera shutters are especially desired in privacy-sensitive locations, such as bathrooms.
先前的研究还调查了应对这些问题的方法，包括技术措施，如实施流量整形技术[5]、自动配置智能设备和实施自动更新[30]，或引入框架，根据情境[41]或预先定义的隐私区域[7]自动调整智能家居的隐私级别。此外，通过共同设计研究，Yao 等人[55]提出了不同的控制机制，如断开设备与互联网的连接并保持本地数据，增加透明度和控制，以及通过不同模式提供访问控制。除这些方法外，最近的研究重点是有形控制机制[3, 14, 38, 52]。这些机制的一个主要优势是易于理解，从而建立信任并确保包容性，尤其是对技术理解能力较低的人而言[3, 52]。此外，Chalhoub 等人[12] 发现，在浴室等对隐私敏感的场所，人们尤其希望有实体摄像头快门。

Sensitive data collected in homes can be exploited, raising various privacy concerns. Yet, traditionally, smart devices were static and had limited interaction capabilities. Future smart assistants will have advanced capabilities through advancements in AI and robotics, enabling completely new ways to invade privacy. Hence, we must understand how such increased capabilities affect users' privacy in home contexts.
家庭中收集的敏感数据可能会被利用，从而引发各种隐私问题。然而，传统的智能设备是静态的，交互能力有限。通过人工智能和机器人技术的进步，未来的智能助手将拥有更先进的功能，从而以全新的方式侵犯隐私。因此，我们必须了解这些功能的增强会如何影响用户在家庭环境中的隐私。

Domestic robots have advanced locomotion and interaction capabilities, enabling them to access all private spaces. This means that their presence might affect not only informational privacy but also physical, psychological, and social privacy [32]. Many domestic robots are, for example, equipped with mobile cameras, enabling them to take images of users or even children in locations such as the bedroom and bathroom, collect spatial information, or witness conversations unnoticed by the users [10, 15, 46]. Moreover, their verbal communication abilities, often paired with a humanoid appearance, lead to people deliberately sharing sensitive information [32, 48].
家用机器人具有先进的运动和交互能力，能够进入所有私人空间。这意味着它们的存在不仅可能影响信息隐私，还可能影响身体、心理和社会隐私[32]。例如，许多家用机器人都配备了移动摄像头，使它们能够拍摄用户甚至儿童在卧室和浴室等场所的图像，收集空间信息，或在用户不注意的情况下见证对话[10, 15, 46]。此外，它们的语言交流能力往往与人形外观相匹配，导致人们故意分享敏感信息[32, 48]。

Even though prior research emphasized the dangers caused by the robots' mobility and physicality [11], users are more concerned about the institutional aspects of their privacy [31], such as how manufacturers handle their data and tended to underestimate the impact of domestic robots on their physical privacy. Yet, users report concerns about the robot being misused for malicious purposes, such as stalking or hacking [31]. Moreover, users in an interview study by Lee et al. [28] reported not being concerned about the robot recording their interactions as long as they were aware of it. However, the interviewees were concerned about accidental recordings that might happen while the robot moves or interacts with other entities. Overall, participants agreed they wanted to be notified about such accidental recordings. The authors also found that participants underestimated the robot's capabilities due to its humanoid shape, which led them to believe that the camera was functioning like human eyes and could not see objects behind its back. Hence, they conclude that users must be
尽管之前的研究强调了机器人的移动性和物理性所带来的危险[11]，但用户更关心其隐私的制度方面[31]，例如制造商如何处理他们的数据，并倾向于低估家用机器人对其物理隐私的影响。然而，用户也担心机器人会被滥用于恶意目的，如跟踪或黑客攻击[31]。此外，Lee 等人的一项访谈研究[28]显示，用户只要知道机器人会记录他们的互动，就不会担心。不过，受访者担心机器人移动或与其他实体互动时可能会意外录音。总的来说，受访者都同意他们希望能收到有关此类意外录音的通知。作者还发现，由于机器人的人形外形，受访者低估了机器人的能力，这导致他们认为摄像头的功能与人眼类似，无法看到机器人背后的物体。因此，他们得出结论，用户必须
thoroughly informed about the robots' exact capabilities [28].
彻底了解机器人的确切能力 [28]。

Experts demand that robots actively communicate when they surveil specific areas [32]. Especially only giving a onetime notice upon purchase is not enough; Instead, robots should give dynamic feedback to regularly communicate their privacy state to users [24]. Lutz et al. [32] conducted expert interviews to elicit privacy mitigation strategies for robots Their approaches include being able to switch off a robot, limiting its movement space, employing data anonymization, or even designing the robot's humanoid features (i.e., its eyes and ears) in a way to signal if data is being collected.
专家们要求机器人在监视特定区域时主动进行交流 [32]。特别是仅在购买时一次性通知是不够的；相反，机器人应提供动态反馈，定期向用户通报其隐私状态 [24]。Lutz 等人[32]进行了专家访谈，以了解机器人的隐私保护策略。他们的方法包括关闭机器人、限制机器人的活动空间、采用数据匿名化，甚至设计机器人的仿人特征（即眼睛和耳朵），以便发出是否正在收集数据的信号。

Domestic robots raise various novel privacy concerns. Thus, experts demand that they regularly communicate their privacy states. Yet, we currently lack a systematic understanding of what communication patterns domestic robots can employ and we do not know which patterns perform best regarding measurements such as understandability and trust.
家用机器人引发了各种新的隐私问题。因此，专家们要求它们定期交流自己的隐私状态。然而，我们目前对家用机器人可以采用哪些交流模式缺乏系统性的了解，也不知道哪种模式在可理解性和信任度等方面表现最佳。

We investigate how locomotion and interaction influence users' privacy concerns and how future domestic robots can effectively communicate the state of their privacy-relevant capabilities through the following three research questions:
我们通过以下三个研究问题来探讨运动和交互如何影响用户对隐私的关注，以及未来的家用机器人如何有效地传达其与隐私相关的能力状态：

RQ1. Prior research showed that current smart devices cause various privacy risks [4, 40, 42], making users concerned about their privacy [12, 26, 27, 50]. Yet, current smart home devices are static or have limited interaction capabilities. In contrast, future domestic robots will have increased capabilities, making them even more invasive. Prior research already showed that domestic robots introduce a new range of risks and concerns [11, 28, 46], yet we do not know how the different levels of interaction and locomotion capabilities impact user concerns. Therefore, we ask in our first research question (RQ1): How do privacy concerns change with increasing levels of locomotion and interaction capabilities?
问题 1.先前的研究表明，目前的智能设备会造成各种隐私风险 [4, 40, 42]，使用户担心自己的隐私 [12, 26, 27, 50]。然而，目前的智能家居设备都是静态的，或者交互能力有限。相比之下，未来的家用机器人将拥有更强的功能，使其更具侵入性。先前的研究已经表明，家用机器人会带来一系列新的风险和担忧[11, 28, 46]，但我们还不知道不同程度的交互和运动能力会如何影响用户的担忧。因此，我们提出了第一个研究问题（RQ1）：随着运动和交互能力水平的提高，隐私问题会发生怎样的变化？

RQ2. Prior research points to the additional risks posed by domestic robots, such as being able to follow us around [46], enter all areas [11], or even make accidental recordings [28]. In response, experts call for domestic robots to communicate their privacy-relevant states to the user regularly [24, 32]. However, research in this regard is scarce. Hence, we ask in our second research question (RQ2): Which patterns should domestic robots employ to communicate their privacyrelevant functionalities to users?
问题 2.先前的研究指出，家用机器人会带来额外的风险，例如可以跟踪我们[46]，进入所有区域[11]，甚至意外录音[28]。为此，专家们呼吁家用机器人定期向用户通报其隐私相关状态 [24, 32]。然而，这方面的研究还很少。因此，我们提出了第二个研究问题（RQ2）：家用机器人应采用哪些模式向用户传达其隐私相关功能？

RQ3. Finally, we need to find out which patterns perform best. In detail, we want to find out which patterns users trust most, which they felt to increase their privacy, which they found most understandable, which they believed to have the best notification qualities, and which they would prefer their smart assistant to use. Hence, we ask in our last research question (RQ3): Which communication patterns perform best regarding trust, privacy, understandability, notification qualities, and general user preference?
问题 3.最后，我们需要找出哪些模式表现最佳。具体来说，我们想知道用户最信任哪种模式，他们认为哪种模式能增加他们的隐私，他们认为哪种模式最容易理解，他们认为哪种模式的通知质量最好，以及他们更希望智能助手使用哪种模式。因此，我们在最后一个研究问题（RQ3）中提出了以下问题在信任度、隐私性、易理解性、通知质量和一般用户偏好方面，哪种通信模式表现最好？

We first set out to understand how increased locomotion and interaction capabilities influence users' privacy concerns in the context of domestic robots. While prior work points to the risks introduced by domestic robots' increased capabilities [11, 28, 46], research on users' concrete concerns is scarce or even shows that users underestimate the impact of robots on their physical privacy [31]. Hence, we conducted an online survey using Prolific to answer our first research question (RQ1). We acquired ethics approval for the survey.
我们首先要了解的是，在家用机器人的背景下，运动和交互能力的增强如何影响用户对隐私的关注。虽然先前的研究指出了家用机器人能力增强所带来的风险[11, 28, 46]，但有关用户具体关注点的研究却很少，甚至表明用户低估了机器人对其身体隐私的影响[31]。因此，我们使用 Prolific 进行了一次在线调查，以回答第一个研究问题（RQ1）。我们获得了调查的伦理批准。

As prior work showed that a multitude of different factors, such as the sensors [35, 50], device manufacturers [36, 56], perceived device utility [56], and familiarity influence users' privacy concerns, we focused on creating descriptions for the various smart assistants with as few biasing factors as possible. Therefore, we used sole textual descriptions and refrained from using pictures or illustrations to not create associations with existing smart home devices or specific manufacturers; relying solely on text is an approach also followed by related work when capturing perceptions of future scenarios [49]. Furthermore, we aligned all texts and only varied the locomotion and interaction capabilities descriptions. Four researchers, two with expertise in privacy and two in human-robot interaction, collaboratively created the different interaction and locomotion stages by clustering the most popular smart assistants according to their capabilities and extending them with the full human-like capabilities, world movement and full interaction to represent future smart assistants. This process resulted in three interaction stages and four locomotion stages, which we combined to create descriptions for 12 smart assistants. All descriptions used the following structure: "Imagine the following
先前的研究表明，传感器[35, 50]、设备制造商[36, 56]、感知设备效用[56]和熟悉程度 等多种不同因素会影响用户对隐私的关注，因此我们将重点放在为各种智能助手创建描述上，尽可能减少偏差因素。因此，我们只使用文字描述，不使用图片或插图，以免用户对现有智能家居设备或特定制造商产生联想；仅依靠文字描述也是相关工作在捕捉未来场景感知时采用的方法[49]。此外，我们对所有文本进行了统一，仅对运动和交互能力的描述进行了改动。四位研究人员（其中两位在隐私方面具有专长，两位在人机交互方面具有专长）共同创建了不同的交互和运动阶段，方法是根据最流行的智能助手的功能对其进行分组，并将其扩展为完全类人功能、世界运动和完全交互功能，以代表未来的智能助手。这一过程产生了三个交互阶段和四个运动阶段，我们将它们结合起来，为 12 个智能助手创建了描述。所有描述均采用以下结构："想象一下

scenario - You own a smart assistant that you
情景--您拥有一个智能助手，您可以

are using in your home. It has the following
在家中使用。它具有以下功能

capabilities: [Locomotion Capability] + The
能力：[运动能力] + The

smart assistant possesses sensing abilities
智能助理具备感知能力

that enable it to comprehend its surroundings + [Interaction Capability]." We revised these textual descriptions through several rounds of discussions before we conducted pilot tests with two researchers in the field of human-computer interaction who were not involved in this project and with 10 participants from Prolific. In response to piloting, we made the locomotion capability descriptions more comprehensive. This resulted in the following texts:
使其能够理解周围环境 + [交互能力]"。在与两位未参与本项目的人机交互领域研究人员以及 Prolific 的 10 位参与者进行试点测试之前，我们通过多轮讨论对这些文字描述进行了修订。根据试点测试结果，我们对运动能力的描述更加全面。最终形成了以下文本：

Locomotion Capabilities. Stationary: The smart assistant is stationary, which means it is constrained to the exact position where you placed it. Linear Movement: The smart assistant can move along a defined path, meaning its movement is constrained by the path you defined. Planar Movement: The smart assistant can move freely around flat, even surfaces,
运动能力。静止：智能助手是静止的，这意味着它被限制在你放置它的准确位置。直线运动：智能助手可以沿着定义的路径移动，这意味着它的移动受您定义的路径限制。平面移动：智能助手可以在平坦、平整的表面上自由移动、
which means that it can freely move around all accessible areas as long as they are on the same floor. World Movement: The smart assistant can move freely across all areas, which means it can move around all accessible areas, even if they are not on the same floor.
这意味着，只要在同一楼层，它就可以在所有可进入的区域自由移动。世界移动：智能助理可以在所有区域内自由移动，这意味着它可以在所有可进入的区域内移动，即使这些区域不在同一楼层。

Interaction Capabilities. Passive Interaction: Yet, the smart assistant can not physically manipulate the environment, objects, or itself. This implies it can perceive individuals and objects within its field of view and analyze associated information. Limited Interaction: While the smart assistant can automatically adjust its orientation to observe its full surroundings, it can not physically manipulate the environment or objects. This implies it can perceive individuals and objects and analyze associated information. Full Interaction: The smart assistant can automatically adjust its orientation to observe its full surroundings and physically manipulate the environment, objects, and itself. This implies it can perceive individuals and objects and analyze associated information.
交互能力。被动交互：然而，智能助理无法实际操作环境、物体或自身。这意味着它可以感知视野内的个人和物体，并分析相关信息。有限交互：虽然智能助理可以自动调整方向以观察周围的所有环境，但它无法实际操控环境或物体。这意味着它可以感知个人和物体并分析相关信息。完全交互：智能助理可以自动调整方向，观察周围的所有环境，并对环境、物体和自身进行物理操作。这意味着它可以感知个人和物体，并分析相关信息。

We started the survey with demographic questions, used the IUIPC questionnaire [33] to understand participants' general privacy perception, and the ATI questionnaire [18] to understand the sample's technical affinity. Afterward, we confronted participants with all 12 smart assistants in random order. After each smart assistant, we asked the participant to respond to "I am strongly concerned about my privacy due to the presence of the smart assistant" on a 100 -point slider ranging from strongly disagree to strongly agree. We used a visual analog scale (VAS) without ticks to prevent the responses from converging around the ticks, cf. [37]. Moreover, we decided to use VAS, as they have been shown to lead to more precise responses and higher data quality [19]. Finally, as VAS collect continuous data, they allow for more statistical tests [43]. In line with recommendations for scale development, we phrased the statements strongly as mildly phrased statements have shown to result in too much agreement [16].
我们从人口统计学问题开始调查，使用 IUIPC 问卷[33]了解参与者的一般隐私感知，并使用 ATI 问卷[18]了解样本的技术亲和力。之后，我们按照随机顺序让参与者面对所有 12 个智能助手。每使用完一个智能助手后，我们都会要求受试者回答 "由于智能助手的存在，我非常担心自己的隐私"，受试者的回答分为 100 分，从 "非常不同意 "到 "非常同意 "不等。我们使用的是不带刻度的视觉模拟量表（VAS），以防止受试者的回答集中在刻度周围，参见文献[37]。此外，我们还决定使用 VAS，因为事实证明 VAS 能带来更精确的回答和更高的数据质量[19]。最后，由于 VAS 收集的是连续数据，因此可以进行更多的统计测试[43]。根据量表编制建议，我们对陈述进行了强烈措辞，因为温和措辞的陈述已被证明会导致过多的一致性[16]。

Additionally, we asked participants to explain their ratings using free text. To ensure the quality of our data, we saved a timestamp after each section and used an attention check item that randomly asked to either set a slider all the way to the right or the left. For the full questionnaire, see Sec. A.1.
此外，我们还要求参与者用自由文本解释他们的评分。为确保数据质量，我们在每个部分后都保存了一个时间戳，并使用了一个注意力检查项目，随机要求将滑块向右或向左滑动。问卷全文见 A.1 节。

We recruited 151 participants via Prolific. We did not use any reputational filters, and our sample had a mean of 337 approved tasks ( ). We had to exclude 61 participants for (1) giving low-effort responses ( ), meaning they explained their ratings with only 2-4 words (e.g., "NA," "i trust") or copied the same response in all 12 conditions, (2) straight-lining, i.e., consistently rating all conditions with 0 or , (3) failing our attention check (see Sec. A.1, question 4 c entering mismatched demographics between Prolific and our survey ( ), and (5) completing the survey three standard deviations faster than the mean .
我们通过 Prolific 招募了 151 名参与者。我们没有使用任何声誉筛选器，我们的样本平均有 337 个认可的任务 ( )。我们不得不排除 61 名参与者，因为他们（1）给出了低强度的回答（ ），这意味着他们仅用 2-4 个字（如 "NA"、"i trust"）来解释自己的评分，或者在所有 12 个条件中都复制了相同的回答；（2）直接排队，即在所有条件中都打 0 分，或者在所有条件中都打 0 分、 ；(3) 未通过我们的注意力检查（请参阅章节 A.1，问题 4 c ；(4) 在 Prolific 和我们的调查中输入不匹配的人口统计数据（ ）；(5) 完成调查的速度比平均速度快三个标准差 。

Figure 1: Participants' mean privacy concern over all locomotion and interaction capabilities with boxplots. The trendline represents the change in relation to the locomotion capability.
图 1：参与者对所有运动能力和交互能力的平均隐私关注度方框图。趋势线表示与运动能力相关的变化。

The final 90 participants ( 47 male, 42 female, and one preferred not to disclose) were between 19 and 62 years old ( ). They were located on three continents (Europe, America, and Africa). Most participants (8) lived in Poland, the United Kingdom, and Italy, followed by Spain (7), South Africa (7), and Portugal (6). Among the participants, 72 were employed full-time, 13 were employed part-time, and five were not in paid work. Moreover, 17 participants were students. Our participants' technical affinity according to the ATI scale [18] was measured on a 6 -point scale. We employed the IUIPC questionnaire [33] using a 7 -point Likert scale to understand their general perception of privacy. The results revealed an average rating of for Awareness, for Control, and for Collection. These scores indicate a relatively high level of privacy concerns, cf. [22]. The survey took , and they were compensated with .
最终的 90 名参与者（47 名男性，42 名女性，1 名不愿透露姓名）的年龄在 19 岁到 62 岁之间（ ）。他们分布在三大洲（欧洲、美洲和非洲）。大多数参与者（8 人）居住在波兰、英国和意大利，其次是西班牙（7 人）、南非（7 人）和葡萄牙（6 人）。参与者中有 72 人从事全职工作，13 人从事兼职工作，5 人没有从事有偿工作。此外，还有 17 名参与者是学生。根据 ATI 量表[18]，我们对参与者的技术亲和力进行了 测量，分值为 6 分。我们采用了 IUIPC 问卷[33]，使用 7 点李克特量表来了解他们对隐私的总体看法。结果显示，对 "意识 "的平均评分为 ，对 "控制 "的平均评分为 ，对 "收集 "的平均评分为 。这些分数表明用户对隐私的关注程度相对较高，参见 [22]。调查采取了 ，他们得到了 的补偿。

We used Python and R to analyze our quantitative data and affinity diagramming [21] for the qualitative data. Here, we printed all statements so two researchers could collaboratively extract the themes by grouping them. We then created headers for each group, frequently rearranged the items, and refined the themes through multiple discussion rounds.
我们使用 Python 和 R 来分析定量数据，使用 affinity diagramming [21] 来分析定性数据。在这里，我们将所有陈述打印出来，这样两名研究人员就可以通过分组合作提取主题。然后，我们为每组创建标题，经常重新排列项目，并通过多轮讨论完善主题。

As our data were not normally distributed ( .001), we used an ART ANOVA [54], which revealed significant effects for LocOMOtion and InTERACTION ( ) ( ) while indicating no interaction effect ( ), see Fig. 1. Pairwise post hoc tests using Wilcoxon signed rank tests with Holm-Bonferroni corrections applied showed that the Locomotions are rated significantly different (linear stationary , and all others ). Moreover, all InteractionS were rated
由于数据不呈正态分布 ( .001)，我们使用了 ART 方差分析 [54]，结果显示 LocOMOtion 和 InTERACTION ( ) ( ) 有显著影响，但没有交互影响 ( ) ，见图 1。使用 Wilcoxon 符号秩检验和 Holm-Bonferroni 校正进行的配对事后检验表明，Locomotions 被评为显著不同（线性 静止 ，所有其他 ）。此外，所有交互作用都被评为
significantly different (passive limited , and all others for all ). We assumed an equidistant distribution between the smart assistants and fitted a line to all mean concern ratings, see Fig. 1. As all trendlines are positive, we conclude that higher locomotion freedom and more interaction capabilities lead to greater privacy concerns.
明显不同（被动 限制了 ，而其他所有 ）。我们假设智能助手之间存在等距分布，并将所有平均关注度评级拟合成一条直线，见图 1。由于所有趋势线均为正值，我们得出结论：更高的运动自由度和更强的交互能力会导致更多的隐私问题。

From the free text descriptions of the participants, we formulated three themes: Concerns Rooted in Locomotion, Concerns Rooted in Interaction, and Additional User Concerns.
根据参与者的自由文本描述，我们拟定了三个主题：基于运动的关注点、基于交互的关注点以及其他用户关注点。

We report our participants' explanations of how the different LOCOMOTION capabilities influence their privacy concerns.
我们报告了参与者对 LOCOMOTION 的不同功能如何影响其隐私问题的解释。

Stationary. Our participants felt most in control over what the assistant could hear and see in the stationary condition. P31, for example, explains that they "would try to place it in a space where no personal activities or situations [are] accessible." Such a non-concerning space could be the kitchen, where the participants do not expect personal conversations to occur but consider the smart assistant especially useful for playing music or providing recipes ( P 43 ).
静止。在静止状态下，我们的参与者对助手所能听到和看到的内容最有控制感。例如，P31 解释说，他们 "会尽量把它放在一个没有个人活动或情况的空间里"。这样的非关注空间可以是厨房，参与者不希望在那里发生私人对话，但认为智能助手在播放音乐或提供食谱时特别有用（P43）。

Linear. Our participants explained that the linear movement would reduce their concerns as they can specify the areas the assistant can access. P53, for example, states: "Because the path is pre-defined, [...] I'd simply avoid putting the smart assistant in the rooms I would like to have privacy in." P29 further states that they would redefine the assistant's path should their preferences or concerns change.
线性。与会者解释说，线性移动可以减少他们的顾虑，因为他们可以指定助手可以进入的区域。例如，P53 表示"因为路径是预设的，[......]我只需避免把智能助理放在我希望有隐私的房间里。P29 进一步指出，如果他们的喜好或关注点发生变化，他们会重新定义助手的路径。

Planar. In contrast to the two more restricted movement capabilities, the planar movement increased our participants' privacy concerns significantly, as P14 explains: "If the assistant is left to roam free, it can collect information at will, and that is a clear security and privacy concern." Yet, participants still felt the assistant's inability to climb stairs or move to different floors helped in preserving some privacy: "Due to its limitation to one floor I might feel a bit safer with my privacy, I can move downstairs or upstairs" (P43).
平面。正如 P14 解释的那样，与两种更受限制的移动功能相比，平面移动大大增加了参与者对隐私的担忧："如果让助手自由行动，它就可以随意收集信息，这显然是一个安全和隐私问题"。然而，参与者仍然认为，助手不能爬楼梯或移动到不同的楼层有助于保护一些隐私："由于它只能在一个楼层活动，我可以下楼或上楼，我可能会觉得自己的隐私更安全一些" （P43）。

World. Our participants were most concerned in the world movement condition as they feared the smart assistant could follow them everywhere, leaving no protected space: "Being able to move even to different floors means there is no safe place in the house" (P83). Moreover, participants were concerned about the assistant showing up unexpected (P56): "It's hard to avoid it popping up unexpectedly, isn't it?"
世界。在 "世界移动 "条件下，我们的参与者最为担忧，因为他们担心智能助理会跟着他们到处走，不给他们留下任何受保护的空间："甚至可以移动到不同的楼层，这意味着家里没有安全的地方"（P83）。此外，参与者还担心助手会突然出现（P56）："很难避免它突然出现，不是吗？

We now report the influence of the different INTERACTION capabilities on participants' privacy concerns.
我们现在报告不同的 INTERACTION 功能对参与者隐私关注的影响。

Passive Interaction. In the passive condition, most participant responses again revolved around the notion of control.
被动互动。在被动条件下，大多数参与者的回答还是围绕着控制的概念。
Participants felt less concerned about their privacy, as they would have "full control on what it sees" (P66), and P70 mentioned that the assistant could " only see what I want." Here, familiarity also played a role as participants knew stationary smart assistants from their daily life, as P83 states: "That's the standard setup of intelligent assistant, so no concern."
参与者对自己的隐私不太担心，因为他们可以 "完全控制它所看到的内容"（P66），P70 提到助手可以 "只看我想看的内容"。在这里，熟悉感也起到了一定的作用，因为参与者在日常生活中就认识固定式智能助手，如 P83 所说：正如 P83 所说："这就是智能助手的标准配置，所以不用担心"。

Limited Interaction. In contrast, the limited interaction capability made our participants way more concerned. Here, P27, for example, compared such a smart assistant to a big brother's eye that would follow them around. In addition, due to its new capabilities, our participants felt less in control over what the smart assistant could perceive: "It can adjust its sight to some parts I do not want to" (P75).
互动有限。相比之下，有限的交互能力让我们的参与者更为关注。例如，P27 把这种智能助手比作会跟着他们到处跑的大哥的眼睛。此外，由于它的新功能，我们的参与者感觉对智能助手所能感知的东西的控制能力更弱了："它可以将视线调整到一些我不想看到的地方"（P75）。

Full Interaction. In addition to the concerns reported regarding the passive and limited interaction capabilities, our participants were now also concerned about the assistant entering all spaces, leaving virtually no room for privacy. As the robot could now "probably open doors and enter areas in times where [I] don't want [it] to" (P43). Additionally to this concern, participants also reported a sense of unease thinking about how the assistant could physically "search the data it wants" (P1) by searching through personal belongings (P30).
全面互动。除了对被动和有限交互能力的担忧之外，我们的参与者现在还担心机器人助手会进入所有空间，几乎没有隐私空间。因为机器人现在 "可能会在我不想让它进入的时候打开门并进入某些区域"（P43）。除了这种担心之外，与会者还表示，想到机器人助手可以通过搜索个人物品来 "搜索它想要的数据"（P1）（P30），他们就感到不安。

Our participants also reported additional concerns not rooted in the robot's interaction and locomotion capabilities. The smart assistant's internet connectivity was the most commonly mentioned concern . Here, participants were concerned that the smart device might share their data, either with the device manufacturer or third parties. For example, P17 stated that they are "always concerned about the type of data [smart devices] can provide to their creator" and P46 said that they would "question if the assistant passes what it perceives to a third party or a remote server." As a possible remedy, P 43 suggested having an offline assistant or one that can only connect to specific applications. The second most common concern was the assistant's video camera sensor, as P57 stated: "I don't like to be watched." P1 was especially concerned about being filmed in intimate situations: "They can probably see me naked while I leave the bathroom." This concern was followed by the audio sensor, which 11 participants mentioned. P27, for example, was concerned that the assistant "might be recording conversations", and P43 mentioned that they would even be concerned about the stationary assistant having good enough microphones to eavesdrop on conversations that might be happening in a different room. Moreover, ten participants mentioned being concerned about the assistant getting hacked, giving criminals access to their sensitive data. P52, for example, wrote: "Someone could hack onto it and know how my home is "built" and break into it." Finally, eight participants were concerned about the assistant storing data: "I do not know where the data is saved" (P69). Less commonly mentioned were concerns regarding the -
我们的参与者还报告了与机器人的交互和运动能力无关的其他问题。 最常提到的顾虑是智能助手的互联网连接。在这方面，参与者担心智能设备可能会与设备制造商或第三方共享他们的数据。例如，P17 表示，他们 "总是担心 [智能设备] 能够向其创建者提供的数据类型"，P46 表示，他们会 "质疑助手是否会将其感知到的信息传递给第三方或远程服务器"。作为一种可能的补救措施，P43 建议使用离线助手或只能连接到特定应用程序的助手。第二个最常见的 问题是助手的摄像头传感器，正如 P57 所说："我不喜欢被人监视。P1 尤其担心在亲密场合被拍摄："我离开浴室时，他们可能会看到我的裸体。紧随其后的是音频传感器，有 11 位与会者提到了这一点。例如，P27 担心助手 "可能会记录谈话内容"，P43 提到他们甚至担心固定助手有足够好的麦克风来窃听可能在不同房间发生的谈话。此外，有 10 位与会者提到担心助手被黑客攻击，让犯罪分子获取他们的敏感数据。例如，P52 写道："有人可能会入侵它，知道我家是如何 "建造 "的，然后闯入我家。最后，8 位与会者对助手存储数据表示担忧："我不知道数据保存在哪里"（P69）。较少提及的是对 -的担忧。
tection of activity data and identification .
保护活动数据 和识别 。

We focus the remainder of this paper on clearly communicating the state of the capabilities our participants most frequently mentioned: internet connectivity, cameras, and audio sensors. Yet, it is important to note that concerns go beyond the pure collection of data, e.g., what could be inferred from the collected data. Yet, to clearly define the scope of this paper, we leave such investigations to future work.
本文的其余部分将重点阐述与会者最常提及的能力状况：互联网连接、摄像头和音频传感器。然而，需要注意的是，我们关注的不仅仅是单纯的数据收集，例如，从收集到的数据中可以推断出什么。然而，为了明确界定本文的研究范围，我们将此类研究留待今后的工作中进行。

While prior research demanded that domestic robots clearly communicate their current privacy state to users [24, 32], research on concrete communication patterns is lacking. Hence, we ran three focus groups with 22 participants to answer (RQ2). We used focus groups to join diverse perspectives and spark creativity. Our ethics committee approved the study.
虽然之前的研究要求家用机器人向用户明确告知其当前的隐私状态[24, 32]，但缺乏对具体沟通模式的研究。因此，我们开展了三个有 22 人参加的焦点小组来回答（问题 2）。我们利用焦点小组来加入不同的观点并激发创造力。我们的伦理委员会批准了这项研究。

We asked participants for their informed consent and demographics. We continued with an introductory round and prior experiences with smart homes and robotic systems. Next, we presented a variety of smart home assistants using pictures and short video clips, aiming to portray the diverse landscape of capabilities and shapes. We started with stationary devices without interaction capabilities and ended with humanoid robots with world movement and full interaction capabilities. As most participants had little experience with robotic systems, it was important to show the diversity to elicit a set of patterns applicable to various domestic robots. Next, we focused on the sensing capabilities of domestic robots, ensuring that they knew that the robots were not restricted to a camera and microphone placed visibly in the front but that the sensing units could be placed everywhere. We then split them into pairs to discuss the risks introduced by domestic robots.
我们询问了参与者的知情同意和人口统计数据。接着，我们向参与者介绍了智能家居和机器人系统。接下来，我们用图片和视频短片展示了各种智能家居助手，旨在描绘出功能和外形的多样性。我们从不具交互功能的固定设备开始，到具有全球移动和全面交互功能的仿人机器人结束。由于大多数参与者对机器人系统的使用经验很少，因此展示机器人的多样性以引出一套适用于各种家用机器人的模式非常重要。接下来，我们重点讨论了家用机器人的传感能力，确保他们知道机器人并不局限于在前方明显位置放置摄像头和麦克风，传感装置可以放置在任何地方。然后，我们将他们分成两组，讨论家用机器人带来的风险。

Next, we presented two privacy-relevant future scenarios with domestic robots. In the first scenario, a person sat at the kitchen table, reviewing medical files while discussing the results with their doctor. In the second scenario, a person was getting ready in the bathroom while ranting about their day. We included a domestic robot in both scenarios to make clear that there are scenarios where robots can help with chores but where we also require privacy. Next, we discussed how current smart assistants communicate their privacy state, showing the Alexa Show's camera shutter and the Amazon Echo's microphone-mute button. We contrasted this with how humans communicate that they are not listening or watching
接下来，我们介绍了两个与家用机器人隐私相关的未来场景。在第一个场景中，一个人坐在厨房的餐桌旁，一边查看医疗档案，一边与医生讨论结果。在第二个场景中，一个人一边在浴室里做准备，一边唠叨着自己一天的生活。我们在这两个场景中都加入了一个家用机器人，以表明在有些场景中，机器人可以帮助人们做家务，但我们也需要隐私。接下来，我们讨论了当前的智能助手如何交流隐私状态，展示了 Alexa Show 的相机快门和亚马逊 Echo 的麦克风静音按钮。我们将其与人类如何表达自己没有在听或看的信息进行了对比。

We introduced the four locomotion stages and the three interaction capabilities. We divided them into pairs and did three rounds of discussions and presentations. For each round, every pair had the same interaction capability: passive interaction, limited interaction, or full interaction. Yet, every pair had a different locomotion capability to join diverse perspectives and animate them to consider their robot's specific skills. We had at least two physical variants of each locomotion and interaction capability in the room to have something graspable for them to interact with. We randomized the order of the interaction capabilities for each focus group to reduce biases. We handed them pen and paper to sketch their ideas. Examples of the sketches can be found in the Appendix Fig. 4. The task was to develop as many communication patterns as possible that signify the state of the camera, microphone, and internet connectivity. We focused on cameras, microphones, and internet connectivity as we found that users were most concerned about them in our first survey. Finally, we had a last group discussion to reflect on the communication patterns invented and to discuss the future of domestic robots in general.
我们介绍了四个运动阶段和三种互动能力。我们将他们分成两人一组，进行了三轮讨论和演示。在每一轮中，每对学生都具有相同的互动能力：被动互动、有限互动或全面互动。然而，每对机器人都有不同的运动能力，以便将不同的观点结合起来，并让他们考虑到自己机器人的特定技能。我们在房间里为每种运动和交互能力都准备了至少两种实体变体，以便让他们能够抓住一些东西进行交互。我们为每个焦点小组随机安排了交互能力的顺序，以减少偏差。我们把纸笔交给他们，让他们画出自己的想法。草图示例见附录图 4。我们的任务是开发出尽可能多的通信模式，以表示摄像头、麦克风和网络连接的状态。我们将重点放在了摄像头、麦克风和网络连接上，因为我们在第一次调查中发现，用户最关心的就是它们。最后，我们进行了最后一次小组讨论，对发明的交流模式进行反思，并对家用机器人的未来进行了总体讨论。

We recruited 22 participants ( 12 male, and 10 female) based on demographics they provided through a pre-screening questionnaire via a university mailing list. They were between 19 and 65 years old with different cultural and educational backgrounds, and came from eight different countries, namely Germany (8), India (5), USA (3), China (2), Brazil (1), South Korea (1), Jordan (1), and Bangladesh (1). They also had different educational backgrounds in computer science (6), biology (3), physics (3), electrical engineering (2), psychology (2), mathematics (2), data science (1), journalism (1), political science (1), and business (1). Their average technical affinity according to the ATI scale [18] was . Six participants had never interacted with a robotic system before, nine 1-3 times, one 4-7 times, and six more than 7 times. They received for the 2 h session.
我们根据参与者通过大学邮件列表提供的预选调查问卷，招募了 22 名参与者（12 名男性和 10 名女性）。他们的年龄在 19 岁到 65 岁之间 ，具有不同的文化和教育背景，分别来自 8 个不同的国家，即德国（8 人）、印度（5 人）、美国（3 人）、中国（2 人）、巴西（1 人）、韩国（1 人）、约旦（1 人）和孟加拉国（1 人）。他们的教育背景也各不相同，包括计算机科学（6 人）、生物学（3 人）、物理学（3 人）、电子工程学（2 人）、心理学（2 人）、数学（2 人）、数据科学（1 人）、新闻学（1 人）、政治学（1 人）和商学（1 人）。根据 ATI 量表[18]，他们的平均技术亲和力为 。6 名参与者从未与机器人系统互动过，9 名 1-3 次，1 名 4-7 次，6 名超过 7 次。他们在 2 小时的课程中获得了 。

We transcribed all focus groups and analyzed the data using thematic analysis [8] and Atlas.ti. First, two researchers independently open-coded the data. They then discussed their codes, resolved ambiguities, and formed code groups. Afterward, a third researcher joined to refine the code groups and extract overarching themes. This process resulted in 202 individual codes, 15 code groups, and six themes. The themes INTERVENTIONS and AWARENESS MECHANISMS form our 86 communication patterns. We also identified the themes TRUST and USABILITY, classifying our patterns further and discussing their applicability. The last theme is HumAnOID vS. NON-HUMANOID, discussing anthropomorphic robots.
我们誊写了所有焦点小组的内容，并使用专题分析 [8] 和 Atlas.ti 对数据进行了分析。首先，两名研究人员独立对数据进行开放式编码。然后，他们讨论了各自的编码，解决了模糊之处，并组成了编码组。之后，第三位研究人员加入进来，完善代码组并提取总体主题。这一过程产生了 202 个单独的代码、15 个代码组和六个主题。主题 "干预 "和 "认识机制 "构成了我们的 86 种交流模式。我们还确定了 "信任 "和 "可用性 "这两个主题，对我们的模式进行了进一步分类，并讨论了它们的适用性。最后一个主题是人类与非人类。非人类，讨论拟人机器人。

This theme consists of all communication patterns that not only signal that a capability is deactivated but physically
这一主题包括所有通信模式，这些模式不仅发出能力停用的信号，而且在物理上
prevent its function. The patterns in this theme can be further divided into physical robot constraints, physical location constraints, and attached props control. Physical robot constraints describes all communication patterns where the robot physically interferes with its capabilities. It ranges from less extreme interventions, such as turning the sensors away ( P 2 , , covering the ears with the hands ( ), or detaching individual sensors ( ), to extreme interventions, such as removing the whole head (P2, P16) or even self-destruction (P10, P12). P13 explains how detaching the sensors could look like: "Having a camera, microphone and a connectivity module and using the hands; basically, the robot taking it off itself, making it very clear that it's not connected." In physical location constraints, our participants discussed interventions that restrict the robot's movement and, thus, its functionalities. Such patterns included the robot blocking its own movement (P2, P5, P10, P12), going to its docking station (P5, P6, P7), or even entering physical confinement (P2, P5, P15, P12, P16, P20, P19), as P15 explains: "[...] a box, like a parking spot, which is like a Faraday box, where no Wi-Fi connection can come through. It's a non-transparent box, and it's soundproof." The last group, attached props control, contains all patterns where the robot has a privacy prop attached, which blocks the robot's functionality. Here, our participants referred to classical camera shutters (P2, P21) but also cables (P5, P11) and switches (P4, P5, P14) that are solely attached to physically interfere with a capability "and when you want to shut it down, just press the switch like a light, and everything will be shut off" (P14).
阻止其功能。本主题中的模式可进一步分为物理机器人限制、物理位置限制和附加道具控制。物理机器人限制描述了机器人对其功能进行物理干预的所有通信模式。它包括一些不太极端的干预措施，如将传感器转开（P 2， ）、用手捂住耳朵（ ）或分离单个传感器（ ），也包括一些极端的干预措施，如移除整个头部（P2，P16）甚至自毁（P10，P12）。P13 解释了如何分离传感器："有一个摄像头、麦克风和一个连接模块，然后用手；基本上，机器人自己把它取下来，非常清楚地表明它没有连接"。在物理位置限制方面，我们的参与者讨论了限制机器人移动，从而限制其功能的干预措施。正如 P15 所解释的那样，这些模式包括机器人阻止自己的移动（P2, P5, P10, P12），进入对接站（P5, P6, P7），甚至进入物理禁闭（P2, P5, P15, P12, P16, P20, P19）：正如 P15 所解释的："[......]一个盒子，就像一个停车位，就像一个法拉第盒子，没有 Wi-Fi 连接可以通过。这是一个不透明的盒子，而且是隔音的"。最后一组 "附加道具控制 "包含了机器人附加隐私道具的所有模式，这些道具会阻挡机器人的功能。在这里，我们的参与者提到了经典的相机快门（P2，P21），但也提到了电缆（P5，P11）和开关（P4，P5，P14），这些附加道具只是为了物理干扰机器人的功能，"当你想关闭它时，只需按下开关，就像按下电灯一样，一切都会被关闭"（P14）。

In contrast to the above theme Interventions, AwareNESS MECHANISMS do not physically prevent a capability but raise users' awareness of the robot's current privacy state This theme consists of the following code groups: Physical robot manipulation, attached props feedback, environment interaction, visual feedback, and audio feedback. Physical robot manipulation contains all the ways a robot can change its own appearance to indicate its current privacy-relevant state, including using hand gestures, such as covering the eyes to signal that it is not watching or crossing the arms to signal the Wifi is disconnected (P20, P22), as P20 explains "you cross your arms out of frustration." Other suggestions included showing empty connectivity ports to the user (P16), retracting sensors (P19), and signaling disengagement through the body posture (P5, P8, P12, P16, P19, P22): "These robots could also just let the arms fall, you can see that the motors and everything are disengaged" (P12). Lastly, the participants also suggested that the robot changes its shape to signal that its capabilities are not activated (P1, P2, P5, 19): "So it could be in a special form when it's active, but while it's deactivated, it could fall into a different form so you know... shape changing" (P19). The group attached props feedback encompasses all patterns where the robot has privacy-specific artifacts attached to communicate the privacy state. This included waving a banner to signal that a capability was deactivated (P7), or attaching a light band (P5), or fake antenna: "Put an antenna or something physical on there that has no use except that it would maybe illuminate red if it's not connected to the internet" (P20). In environment interaction, our participants discussed how the robot could use smart lights installed in the home to communicate its privacy state (P2, P7): "I see a flickering of the light; So it indicates, okay, it's not listening anymore" (P2). In visual feedback, we summarized all traditional patterns requiring a screen or using simple light feedback (P1, P2, P5, P7, P8, P11, P15, P20 - P22). Our participants had diverse ideas of what could be displayed on the screen, ranging from simple text ( ) to symbols ( P 20 , P 22 ), gestures ( P 3 ), and a humanoid face ( to turning the screen off (P2). Lastly, our participants suggested using some form of audio feedback, such as playing distinct sounds (P4, P10, P20, P22) or using the robot's voice (P7, P8, P20, P22): "It says: I'm not listening now" (P20).
与上述主题 "干预 "不同的是，"意识 "机制并不实际阻止机器人的能力，而是提高用户对机器人当前隐私状态的意识：机器人物理操作、附属道具反馈、环境交互、视觉反馈和音频反馈。机器人的物理操作包含机器人改变自身外观以显示其当前隐私相关状态的所有方式，包括使用手势，如遮住眼睛表示没有在看，或交叉双臂表示 Wifi 已断开（P20、P22），正如 P20 所解释的，"你交叉双臂是出于无奈"。其他建议还包括向用户显示空的连接端口（P16）、收回传感器（P19）以及通过身体姿势发出脱离信号（P5、P8、P12、P16、P19、P22）："这些机器人也可以让手臂垂下，你可以看到电机和所有东西都脱离了"（P12）。最后，参与者还建议机器人改变形状，以表示其功能未被激活（P1、P2、P5、19）："因此，当它处于激活状态时，它可能会以一种特殊的形式存在，但当它停止激活时，它可能会变成另一种形式，所以你知道......形状改变"（P19）。小组附加的道具反馈包括机器人附加隐私特定器物以传达隐私状态的所有模式。这包括挥舞横幅以表示某项功能已停用（P7），或附加一个光带（P5），或假天线："在上面放一根天线或其他实物，除了在没有连接到互联网时可能会亮起红灯之外，没有其他用处"（P20）。 在环境互动中，我们的参与者讨论了机器人如何利用安装在家中的智能灯来传达其隐私状态（P2，P7）："我看到灯在闪烁，这表明，好吧，它不再监听了"（P2）。在视觉反馈方面，我们总结了所有需要屏幕或使用简单灯光反馈的传统模式（P1、P2、P5、P7、P8、P11、P15、P20 - P22）。我们的参与者对屏幕上可以显示的内容有不同的想法，从简单的文字（ ）到符号（P 20、P 22）、手势（P 3）和人脸（ 到关闭屏幕（P2）。最后，我们的参与者建议使用某种形式的声音反馈，例如播放不同的声音（P4、P10、P20、P22）或使用机器人的声音（P7、P8、P20、P22）："它说：我现在没在听"（P20）。

This theme describes the factors influencing trust in communication patterns. Here, participants discussed that the type of robot determines their preferred communication patterns. While they considered stationary robots as not very invasive and, thus, requiring less invasive strategies (P5), they discussed that robots with more extreme capabilities also require extreme interventions (P10, P11): "I think that self-destruct is still useful. When your robot has so many capabilities, you also need very strong limitations" (P11). Our participants also discussed that they prefer manual over system control for such invasive robotic systems. That means they preferred mechanisms where the robot can not reactivate its functionalities by itself (P2, P10, P15, P16, P22). P15 suggested hiding the detached sensors from the robot or adding a physical lock so the robot can not free itself: "We thought about a lock from the outside so the robot could close the lid by itself, but then the human could have like a mechanical lock that he or she puts from the outside to be sure that the robot itself can't reopen it." Lastly, our participants also discussed how AWARENESS MECHANISMS require more trust in the robot and its manufacturer than INTERVENTIONS (P5, P10, P11, P13, P15, P16): "It obviously requires some trust in the company that the lights actually state the true status of the device" (P15). In contrast, P13 explained what they like about INTERVENTIONS: "Even if we can't really trust the company - it's a physical barrier."
这一主题描述了影响信任交流模式的因素。在这里，与会者讨论了机器人的类型决定了他们偏好的交流模式。虽然他们认为静止的机器人入侵性不强，因此需要较少的入侵策略（P5），但他们讨论说，具有更极端能力的机器人也需要极端干预（P10，P11）："我认为自毁仍然有用。当你的机器人拥有如此多的能力时，你也需要非常强的限制"（P11）。我们的参与者还讨论说，对于这种侵入式机器人系统，他们更喜欢手动控制，而不是系统控制。也就是说，他们更喜欢机器人无法自行重新激活其功能的机制（P2、P10、P15、P16、P22）。P15 建议将分离的传感器从机器人身上隐藏起来，或者加一把物理锁，这样机器人就无法释放自己了："我们考虑过从外部上锁，这样机器人就可以自己关上盖子，但人类可以从外部上一把机械锁，以确保机器人自己无法重新打开盖子"。最后，我们的参与者还讨论了 "警示机制 "如何比 "干预措施 "更需要机器人及其制造商的信任（P5、P10、P11、P13、P15、P16）："很明显，这需要对公司有一定的信任，相信这些指示灯确实能显示设备的真实状态"（P15）。相反，P13 解释了他们喜欢 INTERVENTIONS 的原因："即使我们不能真正信任公司--这也是一种物理障碍"。

Our participants discussed how the situation influences the applicability of the different patterns and how familiarity, intuitiveness, and joy of use affect their perception of the patterns.
我们的参与者讨论了情境如何影响不同模式的适用性，以及熟悉程度、直观性和使用乐趣如何影响他们对模式的感知。

Our participants discussed, for example, that audio feedback is most effective when the robot is not in the same room or hidden somewhere (P1, P2, P10): "It should also give some audio feedback. So if it's somewhere under my couch, and I can't see it, I know if it's on or off" (P10). Besides, our participants also discussed that many of the INTERVENTIONS are unsuitable if the robot is currently doing a task (P1, P15, P20): "If you tell it: Just go away! That doesn't work if it's still doing a task" (P20). In addition, our participants often considered the INTERVENTIONS inconvenient; for example, when the microphone, camera, and internet are deactivated, there is hardly any way of restarting the robot (P4, P17). Finally, our participants discussed that familiar communication patterns have the big advantage of being immediately understandable (P19), that humanoid patterns are more understandable due to their intuitiveness (P3, P5, P19), and that they would prefer to use patterns they considered fun to use (P7, P10): "It is fun. Like it's something that is trying to mimic me, but it's not me" (P7).
例如，我们的参与者讨论说，当机器人不在同一个房间或隐藏在某个地方时，声音反馈最为有效（P1、P2、P10）："它还应该提供一些声音反馈。这样，如果它在我的沙发下面，我看不到它，我就能知道它是开着还是关着"（P10）。此外，我们的参与者还讨论到，如果机器人正在执行任务，那么许多干预措施都是不合适的（P1、P15、P20）："如果你告诉它走开！如果它还在执行任务，那就行不通了"（P20）。此外，我们的参与者常常认为干预措施很不方便，例如，当麦克风、摄像头和网络被关闭时，几乎没有任何办法可以重新启动机器人（P4，P17）。最后，我们的参与者讨论说，熟悉的交流模式有一个很大的优势，就是可以立即理解（P19），人形模式因其直观性而更容易理解（P3, P5, P19），而且他们更愿意使用他们认为有趣的模式（P7, P10）："这很有趣。就像它在模仿我，但它不是我"（P7）。

Our participants discussed that humanoid robots provoke human expectations as their shape makes them appear more capable (P1, P2), which also makes them feel less controllable (P6) and sometimes even evokes feelings of unease (P2, P6, P7, P11, P20): "I wouldn't want human-like with skin on it or something, because it would be creepy" (P7). The anthropomorphic appearance also led to people discussing whether the robots would then develop some form of consciousness, evoking feelings of pity (P2, P3, P7): "Maybe you get emotionally attached in a way that you feel sorry for them when they have to do certain tasks [...] it feels like enslav ing" (P2). Yet, other participants completely disagreed and stated that they would never feel sorry for a machine, regardless of its appearance (P10, P13). Moreover, our participants also discussed that the human-like shape might evoke feelings of trust, which can be unjustified as the robot might collect and share sensitive data (P8). Finally, the participants debated that while some communication patterns are already weird if used by a human, for example, staying in the same room but covering the eyes to signal that one is not watching (P4), this would become even stranger if adopted by a robot (P5): "If a robot is covering its eyes I would be like: What's wrong with you? Just turn off your camera, dude!"
我们的参与者讨论说，人形机器人会激起人类的期望，因为它们的外形让它们显得更有能力（P1、P2），这也让它们感觉不那么可控（P6），有时甚至会唤起不安的感觉（P2、P6、P7、P11、P20）："我不希望它像人一样，但又有皮肤什么的，因为那会让人毛骨悚然"（P7）。拟人化的外观还引发了人们对机器人是否会产生某种意识的讨论，从而唤起人们的怜悯之心（P2、P3、P7）："也许你会对机器人产生情感依恋，当它们必须完成某些任务时，你会为它们感到遗憾[......]感觉就像被奴役一样"（P2）。然而，其他与会者则完全不同意这种观点，他们表示，无论机器的外观如何，他们都不会为其感到遗憾（P10，P13）。此外，我们的参与者还讨论说，类似人类的外形可能会唤起人们的信任感，而这种信任感可能是没有道理的，因为机器人可能会收集和分享敏感数据（P8）。最后，与会者争论说，有些交流方式如果是人类使用就已经很奇怪了，例如，待在同一个房间里却捂住眼睛，表示自己没有在看（P4），如果机器人采用这种方式，就会变得更加奇怪（P5）：如果机器人遮住眼睛，我会说："你怎么了？你怎么了？快把摄像头关掉，老兄！"

To construct the gesture set, we reviewed all individual quotes in the themes Interventions and Awareness MECHANISMS and merged all quotes that described the same communication pattern. We further categorized the remaining quotations by their tackled functionality, i.e., camera, microphone, or internet connectivity. This process resulted in 86 individual communication patterns, 33 INTERVENTIONS and
为了构建手势集，我们对 "干预 "和 "认知机制 "主题下的所有引文进行了审查，并合并了所有描述相同交流模式的引文。我们还根据所处理的功能（即摄像头、麦克风或网络连接）对其余语录进行了进一步分类。通过这一过程，我们得出了 86 种单独的交流模式，其中 33 种为 "干预 "模式，33 种为 "认知机制 "模式。
53 Awareness Mechanisms. Twenty-eight tackled the camera, 27 the microphone, 21 the internet connectivity, and 10 all functionalities simultaneously. Please refer to Tab. 1 for the complete list of all communication patterns.
53 个感知机制。其中，28 项涉及摄像头，27 项涉及麦克风，21 项涉及互联网连接，10 项同时涉及所有功能。请参见表 1。所有交流模式的完整列表请参见表 1。

Via a large-scale online survey, we determined which patterns performed best regarding trust, privacy, understandability, notification quality, and general user preference (RQ3). Our ethics committee approved the survey.
通过大规模在线调查，我们确定了哪些模式在信任度、隐私性、可理解性、通知质量和一般用户偏好方面表现最佳（问题 3）。我们的伦理委员会批准了这项调查。

The survey started with a short introductory text, instructing the participants to immerse themselves in a future situation where they own a domestic robot that supports them with daily chores. The text further stated that the robot uses a communication pattern to show that the user's privacy is protected. After that, every participant saw one of the 86 communication patterns. For InTERVENTIONS, we used the following sentence structure: The domestic robot does [communication pattern] to physically prevent [capability], and for AWARENESs Mechanisms, we used: The domestic robot does [communication pattern] to signal
调查以一段简短的介绍性文字开始，指示参与者将自己置身于一个未来的情境中，即他们拥有一个支持他们处理日常家务的家用机器人。文字进一步指出，机器人使用一种交流模式来表示用户的隐私受到保护。之后，每位参与者都看到了 86 种交流模式中的一种。在 "干预 "部分，我们使用了以下句子结构：对于 "干预措施"，我们使用了以下句子结构："家用机器人通过[交流模式]来实际防止[能力]"：家用机器人使用[交流模式]来发出信号

that [capability] is deactivated. Next, we asked them to rate eight statements on a 100-point scale (from strongly disagree to strongly agree). We used VAS without ticks for the same reasons as previously stated [19, 37, 43]. We asked (1) how well our participants felt their privacy was protected, (2) how much they trusted the capability to be actually deactivated, (3) how effective, (4) intrusive, (5) noticeable, (6) understandable, and (7) disturbing they considered the communication pattern and finally, (8) how much the participant would like their domestic robot to use the communication pattern. Additionally, we asked them to put the slider all the way to the right side as an attention check. We used the statements of Rzayev et al. [44] to investigate the notification quality (statements (3) to (7)) in line with [51]. For the full questionnaire, see Sec . A.3.
能力]已停用。接下来，我们要求他们用 100 分制（从 "非常不同意 "到 "非常同意"）对八项陈述进行评分。我们使用不带勾的 VAS，原因与之前所述的相同[19, 37, 43]。我们询问：（1）参与者认为他们的隐私得到了多大程度的保护；（2）他们对实际停用功能的信任程度；（3）有效程度；（4）干扰程度；（5）明显程度；（6）可理解程度；（7）他们认为这种交流模式会造成的干扰；最后，（8）参与者希望他们的家用机器人使用这种交流模式的程度。此外，我们还要求他们将滑块完全置于右侧，以检查注意力。我们使用了 Rzayev 等人[44]的陈述来调查通知质量（陈述（3）至（7）），这与[51]一致。调查问卷全文见第 A.3 节。A.3.

We recruited 1720 participants via Prolific as we wanted to have 20 ratings per communication pattern. We used no reputational filters, and our participants had a mean of 490 ( ) approved tasks. We recruited our participants in several batches to (1) replace participants who failed the attention check (see Sec. A.3, question 7) and (2) counterbalance the sample in terms of country of birth and gender. The participants were between 18 and , ) years old, and 869 identified male, 825 as female, 22 as non-binary, and four did not disclose their gender. 1665
我们通过 Prolific 招募了 1720 名参与者，因为我们希望每个交流模式有 20 个评分。我们没有使用声誉筛选器，参与者平均有 490 ( )个通过审核的任务。我们分几批招募参与者，以便：(1) 替换未通过注意力检查的参与者（见第 A.3 节，问题 7） ；(2) 在出生国家和性别方面平衡样本。参与者的年龄在 18 到 , 之间，869 人认为自己是男性，825 人认为自己是女性，22 人认为自己是非二元性别，4 人没有透露自己的性别。1665
were full-time, and 55 were employed part-time, of whom 107 were also students. Most held an undergraduate degree (659), a graduate degree (585), or a high school diploma (208). Our participants were born in 107 different countries. Most had their origin in the UK (123), Poland (102), Portugal (87), Italy (86), South Africa (84), and Mexico (83).We compensated the 1 min survey with .
其中 107 人还是学生。大多数人拥有本科学位（659 人）、研究生学位（585 人）或高中文凭（208 人）。我们的参与者出生在 107 个不同的国家。大多数人出生在英国（123 人）、波兰（102 人）、葡萄牙（87 人）、意大利（86 人）、南非（84 人）和墨西哥（83 人）。

We analyzed our data using Python. First, we employed hierarchical clustering to understand the underlying relationships among the communication patterns. This allowed us to build clusters based on linkage criteria and distance thresholds Thereby, we found three distinct clusters: one consisting of 80 communication patterns, one of five, and one cluster that only contained a single communication pattern. We used principal component analysis (PCA) to reduce the eight measurements (Privacy, Trust, Effectiveness, Intrusiveness, Noticability, Understandability, Disturbance, and Preference) to two dimensions for easier investigation; see Fig. 2. The PCA visualization shows that the big cluster is separated from the two other clusters. To understand the meaning of our clusters, we utilized parallel coordinates plots where we highlighted the separate clusters. Here, Fig. 2a revealed that the two "outlier" clusters comprise all low-scoring communication patterns Five of these "outlier" patterns are represented with a similar curve in the parallel coordinates plot, showing that they scored equally low regarding privacy, trust, and user preference. Those patterns were: (1) the robot covering its ears with its hands (CP10), (2) or facing the wall to prevent audio recordings (CP36), (3) the robot deactivating its rotation function to signal that the camera is off (CP38), (4) the robot facing the wall to signal that the microphone is off (CP38), and (5) the robot parking against a pillow to prevent the microphone from recording (CP46), whereby CP36 and CP46 scored lowest regarding trust and privacy. CP45, the robot killing itself to prevent all capabilities, behaved differently than all other patterns and was perceived as, by far, the most disturbing. Yet, it scored well regarding privacy and trust. We attribute the low scores of these patterns to either their inability to convincingly block a sensor, such as parking against a pillow to interfere with the microphone state, or to the dis connect between the action and targeted capability, such as facing a wall to signal microphone states. Finally, the robot covering its ears might have been perceived as strange or deceptive, and the robot killing itself scored low overall because of its extremely disturbing nature.
我们使用 Python 对数据进行了分析。首先，我们采用了分层聚类来了解交流模式之间的潜在关系。因此，我们发现了三个不同的聚类：一个由 80 种交流模式组成，一个由 5 种交流模式组成，还有一个聚类只包含一种交流模式。我们使用主成分分析法（PCA）将八个测量维度（隐私、信任、有效性、侵扰性、可注意性、可理解性、干扰和偏好）缩减为两个维度，以便于调查；见图 2。PCA 可视化显示，大聚类与其他两个聚类是分开的。为了理解聚类的含义，我们利用平行坐标图来突出显示独立的聚类。图 2a 显示，两个 "离群 "聚类包含了所有得分较低的交流模式，其中五个 "离群 "模式在平行坐标图中以相似的曲线表示，表明它们在隐私、信任和用户偏好方面得分同样较低。这些模式是(1)机器人用手捂住耳朵（CP10），(2)或面向墙壁以防止录音（CP36），(3)机器人关闭旋转功能以表示摄像头已关闭（CP38），(4)机器人面向墙壁以表示麦克风已关闭（CP38），(5)机器人停靠在枕头上以防止麦克风录音（CP46），其中 CP36 和 CP46 在信任和隐私方面得分最低。CP45，即机器人杀死自己以阻止所有功能，其行为与所有其他模式不同，被认为是迄今为止最令人不安的模式。 然而，它在隐私和信任方面的得分却很高。我们将这些模式得分较低的原因归结为它们无法令人信服地阻挡传感器，例如靠着枕头停放以干扰麦克风状态，或者是动作与目标能力之间的联系不紧密，例如面向墙壁以发出麦克风状态信号。最后，机器人捂住自己的耳朵可能会被认为是奇怪或具有欺骗性，而机器人杀死自己的行为则由于其极其令人不安的性质而得分较低。

Moreover, we also highlighted the best-scoring patterns in Fig. 2b. Their opposing position with respect to the lowscoring patterns indicates that the PCA can capture the quality of the patterns. Comparing the insights from both plots, we see that while we found some outliers, most patterns were equally well received. Eight out of the ten best scoring pat- terns are interventions, i.e., actions done by the robot that physically prevent the capability. In detail, the best scoring patterns were: (1) the robot putting a physical cover over its camera (CP51), the robot blocking its own movement (CP6), the robot deactivating its rotation function (CP15), or the robot using a physical switch (CP69) to prevent the camera from recording; the robot removing the microphone's cable (CP56) or detaching the microphone (CP19) to prevent audio recordings; and the robot detaching its memory card (CP1), or going to its docking station (CP39) to prevent all functionalities at once. In contrast, the two best-scoring awareness mechanisms are both human gestures, whereby one was more generally phrased: The robot uses a hand gesture to signal that the microphone is off (CP68), and the other one very concretely: The robot crossing its arm to signal that it is disconnected from the internet (CP14). In summary, most patterns that scored well across all measurements represented interventions that are familiar from the smart home environment (i.e., a camera shutter or going to the docking station) or represent interventions a human would do but applied to the robot (i.e., removing the cable or memory card, detaching the sensor [23]).
此外，我们还在图 2b 中突出显示了得分最高的图案。它们与低分模式的相对位置表明，PCA 可以捕捉到模式的质量。对比这两幅图，我们可以发现，虽然我们发现了一些异常值，但大多数模式的得分都相当高。在 10 个得分最高的模式中，有 8 个是干预模式，即机器人采取的实际阻止能力行动。具体来说，得分最高的模式是(1)机器人在摄像头上盖上物理盖板（CP51）、机器人阻挡自己的移动（CP6）、机器人停用旋转功能（CP15）或机器人使用物理开关（CP69）来阻止摄像头录制；机器人移除麦克风的线缆（CP56）或拆下麦克风（CP19）来阻止音频录制；机器人拆下存储卡（CP1）或进入扩展坞（CP39）来同时阻止所有功能。相比之下，得分最高的两个感知机制都是人类手势，其中一个的措辞更为宽泛：机器人用手势表示麦克风已关闭（CP68），另一个则非常具体：机器人交叉手臂，表示已断开网络连接（CP14）。总之，在所有测量中得分较高的大多数模式都代表了智能家居环境中熟悉的干预措施（如相机快门或前往扩展坞），或者代表了人类会做的干预措施，但却应用在了机器人身上（如移除电缆或存储卡、分离传感器[23]）。

In Fig. 3, we visualize each pattern's average score for the Privacy measurement. Here, we see that the three bestperforming patterns are all interventions, meaning they not only signal the sensor state but physically prevent the functionality. In detail, the three best-performing patterns in regards to Privacy are (1) the robot putting a physical cover over its camera to prevent it from filming (CP51), (2) the robot detaching its microphone (CP19), and (3) the robot removing the microphone's cable (CP56) to prevent audio recordings. In contrast, the three worst-performing patterns are (1) the robot facing the wall to prevent the camera from filming (CP36), (2) the robot covering its ears with its hands to prevent the microphone from functioning ( CP 10 ), and (3) the robot parking against a pillow to prevent audio recordings (CP46). While these patterns are also all interventions, they represent more experimental and unfamiliar patterns. In addition, CP10 has a very large interquartile range (IQR), showing how differently our participants perceived the pattern. Moreover, the rather large IQRs across all communication patterns ( , ) quantify their polarizing nature. We find that seven of the overall best scoring patterns also scored best regarding their mean privacy rating. This shows, on the one hand, the small differences between the patterns and that many scored almost equally well. On the other hand, this shows a high disparity between the measurements, meaning that while a pattern can be perceived as very privacy-preserving, it might not score as well regarding the other measurements, signifying the importance of choosing the right pattern for a specific goal or situation.
在图 3 中，我们展示了每种模式在隐私测量中的平均得分。在这里，我们可以看到表现最好的三种模式都是干预模式，这意味着它们不仅能发出传感器状态信号，还能在物理上阻止功能的实现。具体来说，在隐私方面表现最好的三种模式分别是：（1）机器人在摄像头上盖上物理盖板，防止摄像头被拍摄（CP51）；（2）机器人拆下麦克风（CP19）；（3）机器人拆下麦克风线缆（CP56），防止录音。相比之下，表现最差的三种模式是：(1) 机器人面向墙壁，以防止摄像头拍摄（CP36）；(2) 机器人用手捂住耳朵，以防止麦克风工作（CP10）；(3) 机器人靠在枕头上，以防止录音（CP46）。虽然这些模式也都是干预行为，但它们代表了更多实验性的陌生模式。此外，CP10 的四分位数间距（IQR）非常大，这表明参与者对该模式的感知存在很大差异。此外，所有交流模式（ 、 ）的 IQR 都相当大，这也量化了它们的两极分化性质。我们发现，在总体得分最高的模式中，有七种模式的平均隐私评分也是最高的。这一方面说明模式之间的差异很小，而且许多模式的得分几乎相同。另一方面，这也显示了测量结果之间的巨大差异，也就是说，虽然一种模式可以被认为是非常保护隐私的，但它在其他测量结果中的得分可能并不高，这说明了针对特定目标或情况选择正确模式的重要性。

For an overview of all patterns' means and SDs, see Tab. 1. We created an interactive web app (https://robot-patternsfinder.web.app/) that allows designers and researchers to explore communication patterns based on various requirements.
所有模式的平均值和标差概览见表 1。1.我们创建了一个交互式网络应用程序（https://robot-patternsfinder.web.app/），让设计人员和研究人员可以根据各种要求探索交流模式。

(a) We found three clusters. The communication patterns in the separate clusters performed worse on average than those in the big cluster.
(a) 我们发现了三个集群。不同集群中的通信模式平均表现不如大集群中的通信模式。

Figure 2: Insights into the communication patterns. We reversed the two negative items for semantic readability (R).
图 2：洞察交流模式。我们颠倒了两个负面项目的语义可读性（R）。

We found that domestic robots' increasing locomotion and interaction capabilities lead to heightened privacy concerns (RQ1), that their novel interaction and locomotion capabilities enable new ways to indicate or intervene with their sensor states (RQ2), and that most communication patterns perform equally well, showing that pattern use depends on the specific requirements of a situation (RQ3). In the following, we will discuss and relate our key findings to prior work.
我们发现，家用机器人的运动和交互能力不断增强，这导致了对隐私的更多关注（问题 1）；家用机器人新颖的交互和运动能力带来了利用传感器状态进行指示或干预的新方法（问题 2）；大多数交流模式的表现不相上下，这表明模式的使用取决于具体情况的要求（问题 3）。下面，我们将讨论我们的主要发现，并将其与之前的工作联系起来。

While prior work warned about the privacy threats rooted in domestic robots' increased mobility and physicality [11, 32], there is no work so far linking privacy concerns directly with those capabilities. Quite the contrary, prior work even found that users are only mildly concerned about their physical privacy when dealing with domestic robots [31]. In contrast to this, we found that participants' privacy concerns increase step-wise with rising interaction and locomotion. Our participants explained their increased concerns with loss of control: While, in the case of stationary robots, they could still restrict what the robot could hear and see by placing it in specific areas, robotic systems with various locomotion and interaction capabilities can search through private documents or even unlock doors, leaving virtually no space for privacy.
虽然以前的研究曾警告过家用机器人的移动性和物理性的增强会带来隐私威胁[11, 32]，但迄今为止还没有研究将隐私问题与这些能力直接联系起来。恰恰相反，之前的研究甚至发现，用户在与家用机器人打交道时，对其身体隐私的关注程度很低[31]。与此相反，我们发现，随着交互和运动能力的提高，参与者对隐私的担忧也在逐步增加。我们的参与者用失去控制来解释他们的担忧：在机器人静止不动的情况下，他们还可以通过将其放置在特定区域来限制机器人的听觉和视觉，但具有各种运动和交互能力的机器人系统可以搜索私人文件，甚至可以打开门锁，这就几乎没有了隐私空间。
This was also picked up in the focus groups, where our participants agreed that advanced robot capabilities require stronger communication patterns. Here, our participants suggested awareness mechanisms most frequently for stationary robots with limited interaction capabilities, such as simple light indications or audio feedback. At the same time, they wished for the highest level of privacy when dealing with robots with advanced capabilities. Here, our participants' suggestions most often included intervention mechanisms, but even those were sometimes not perceived as secure enough. As a result, their suggestions also included ways to stop the robot from recovering its functional state, such as adding physical locks to prevent it from leaving a physical enclosure or moving detached sensors and cables out of the robot's reach.
这一点在焦点小组中也得到了体现，我们的参与者一致认为，先进的机器人能力需要更强的交流模式。在这里，我们的参与者为交互能力有限的固定式机器人提出了最常见的感知机制，例如简单的灯光指示或声音反馈。同时，他们也希望在与具有高级功能的机器人打交道时，能够最大程度地保护隐私。在这方面，参与者的建议通常包括干预机制，但即使是干预机制有时也不够安全。因此，他们的建议还包括阻止机器人恢复其功能状态的方法，例如加装物理锁，防止机器人离开物理外壳，或将分离的传感器和电缆移到机器人触及不到的地方。

Key Finding 1: Advanced Capabilities Require Strong Interventions. The more capable a domestic robot is, the more it threatens users' privacy, evoking the desire for mechanisms that provide the highest levels of certainty and trust.
主要发现 1：先进的功能需要强有力的干预。家用机器人的能力越强，对用户隐私的威胁就越大，这就唤起了人们对提供最高级别确定性和信任的机制的渴望。

Our results show that most of the well-scoring patterns either represent familiar interventions, such as physical covers or entering the docking station, or interventions usually employed by humans to mitigate their concerns, such as unplugging cables [23]. We attribute the high scores of these patterns to their tangibility and familiarity, making it easy for users to
我们的研究结果表明，大多数得分较高的模式要么代表了人们熟悉的干预措施，如物理盖板或进入扩展坞，要么代表了人类通常用来减轻担忧的干预措施，如拔下电缆[23]。我们将这些模式的高分归因于它们的有形性和熟悉性，使用户很容易

Figure 3: Mean ratings for the PRIVACY statement. Interventions are green, and awareness mechanisms are blue.
图 3：隐私声明的平均评分。干预措施为绿色，宣传机制为蓝色。

understand how they work. In fact, prior work emphasized the value of employing tangible mechanisms for higher trust and understandability [3], which ultimately contributes to inclusive privacy [52]. Yet, this relationship between familiarity and trust also works the other way around; some patterns scored low as users felt they might not be effective. For example, preventing audio recordings by facing the wall or parking against a pillow. We attribute the low scores to users being aware of the high sensitivity of current audio sensors that can capture noises even when obstructed. Yet, the advantage of familiarity is not only true for tangible mechanisms. Also, human hand gestures scored well in our third study. This can be explained by discussions from our focus group, where par ticipants praised these gestures for being understandable and intuitive. Key Finding 2: Familiarity with a pattern fosters understandability, trust, and general user preference. Such familiarity can stem from smart devices already having similar mechanisms integrated or from applying knowledge and actions from daily life to the novel robotics space.
了解它们是如何工作的。事实上，先前的工作强调了采用有形机制的价值，以提高信任度和可理解性[3]，这最终有助于实现包容性隐私[52]。然而，这种熟悉度与信任度之间的关系也有反作用；有些模式得分较低，因为用户认为它们可能无效。例如，通过面朝墙壁或靠着枕头停车来防止录音。我们将这些低分归因于用户意识到了当前音频传感器的高灵敏度，即使在被遮挡的情况下也能捕捉到声音。然而，熟悉的优势不仅适用于有形装置。在我们的第三次研究中，人类手势的得分也很高。这可以从焦点小组的讨论中得到解释，小组成员称赞这些手势易懂、直观。主要发现 2：熟悉一种模式有助于提高可理解性、信任度和一般用户偏好。这种熟悉感可能源于智能设备已经集成了类似的机制，也可能源于将日常生活中的知识和操作应用到新型机器人领域。

In contrast to our participants' general preference for humanoid hand gestures, other patterns leveraging human metaphors performed badly, such as the robot covering its ears to prevent audio recordings. Our focus groups can explain this. Here, participants discussed that they would find it even weird for humans to cover their ears to signal that they are not listening instead of simply leaving the room. Hence, a robot replicating such behavior would be even more strange Another reason might be the difference between awareness mechanisms and interventions. While signifying the sensor state using hand gestures might be well understandable and, thus, well received, covering the ears as an intervention mechanism might provoke distrust; users might be skeptical that the gesture prevents the recording capability, especially as the robot's microphones are not necessarily placed in the ear.
与我们的参与者普遍偏爱仿人手势形成鲜明对比的是，其他利用人类隐喻的模式表现不佳，例如机器人捂住耳朵以防止录音。我们的焦点小组可以解释这一点。在这里，参与者讨论说，如果人类捂住耳朵表示不听，而不是直接离开房间，他们会觉得很奇怪。另一个原因可能是感知机制和干预之间的差异。用手势来表示传感器的状态可能很容易理解，因此也很受欢迎，而捂住耳朵作为一种干预机制可能会引起不信任；用户可能会怀疑这种手势会妨碍录音功能，尤其是机器人的麦克风并不一定放在耳朵里。

Our focus group participants also discussed that the robot's shape influences their general perception; they agreed that a humanoid shape makes a robot seem more capable. At the same time, however, they also discussed that a humanoid form makes a robot seem less controllable. Some participants even considered a too-humanoid appearance creepy, linking to the well-recognized uncanny valley effect [45], and discussed that their shape might evoke undesired feelings, such as feeling pity for the robot when it has to complete undesired tasks. In this regard, prior work suggested exploring the value of "honest anthropomorphism," meaning using anthropomorphic features to notify the users of what a robot is actually doing [24]. Our results show that while anthropomorphic patterns can help foster understandability and trust, they are sometimes perceived as creepy or weird. Moreover, we found them to be more suitable for awareness mechanisms than for interventions. Key finding 3: While humanoid shapes and behaviors foster understandability through intuitiveness and familiarity, they can also evoke feelings of unease and even creepiness. Hence, we suggest employing anthropomorphism carefully and align it with the specific situation.
我们的焦点小组参与者还讨论了机器人的外形会影响他们的总体感知；他们一致认为，人形外形会让机器人看起来更有能力。但与此同时，他们也讨论说，人形会让机器人看起来不那么容易控制。一些参与者甚至认为过于人性化的外观会让人毛骨悚然，这与广为人知的不可思议谷效应（uncanny valley effect）有关[45]，他们还讨论说，机器人的外形可能会唤起他们不想要的感觉，比如当机器人不得不完成不想要的任务时，他们会觉得机器人很可怜。在这方面，先前的研究建议探索 "诚实拟人化 "的价值，即使用拟人化特征来告知用户机器人实际上在做什么[24]。我们的研究结果表明，虽然拟人化模式有助于提高可理解性和信任度，但有时也会被认为令人毛骨悚然或怪异。此外，我们发现拟人化图案更适合用于认知机制，而非干预机制。主要发现 3：虽然拟人化的形状和行为可以通过直观性和熟悉感提高可理解性，但它们也会唤起不安感，甚至令人毛骨悚然。因此，我们建议谨慎使用拟人化，并根据具体情况加以调整。

In summary, many factors must be considered when choosing the optimal communication pattern. As discussed previously, the more capable and intrusive a robot is, the stronger the employed interventions should be. Similarly, Windl et al. [53] suggest that preventing a situation from being privacy violating should be preferred (i.e., through interventions) in contrast to using notices (i.e., awareness mechanisms) whenever possible. Yet, they also discuss that the right mechanism strongly depends on the constraints of a situation. This is especially true in the case of domestic robots, as it is often not as easy as unplugging the robot or sending it away. In contrast, the robot most often needs its full capabilities to fulfill the tasks it was purchased for in the first place. Hence, which communication pattern to employ also depends on the robots' task and whether it is currently actively working or not. That means that, even though interventions provide higher levels of trust and certainty, sometimes awareness mechanisms might be the better option. Moreover, while familiar patterns are often perceived as very understandable and trustworthy, and using humanoid metaphors should certainly be considered familiar, their usage must still be carefully considered as they walk a
总之，在选择最佳通信模式时，必须考虑许多因素。如前所述，机器人的能力越强，侵扰性越大，采用的干预措施就应该越有力。同样，Windl 等人[53] 建议，在可能的情况下，应优先考虑防止出现侵犯隐私的情况（即通过干预），而不是使用通知（即感知机制）。不过，他们也讨论了正确的机制在很大程度上取决于情况的限制因素。对于家用机器人而言，情况尤其如此，因为这往往不是拔掉机器人插头或将其送走那么简单。与此相反，机器人往往需要其全部能力来完成购买之初的任务。因此，采用哪种通信模式也取决于机器人的任务以及它目前是否在积极工作。这意味着，尽管干预能提供更高的信任度和确定性，但有时认知机制可能是更好的选择。此外，虽然熟悉的模式通常被认为非常易懂和值得信赖，而且使用仿人隐喻当然也应被视为熟悉的模式，但使用时仍须慎重考虑，因为它们会对机器人的工作产生影响。
fine line between being intuitive and creepy.
直观和令人毛骨悚然之间的细微差别。

The varying individual ratings also reflect this discrepancy and polarizing nature of some communication patterns. While the measurements for privacy, trust, and overall user preference seem to mostly correlate (see Fig. 2), the other measurements do not seem to follow a similar pattern: While a communication pattern might convey high levels of privacy and trust, it might also be perceived as disturbing or barely noticeable. In addition, the high variance speaks for a generally highly subjective perception of some patterns. As we recognized this discrepancy between the different measurements and that the importance of individual measurements depends on the characteristics of a situation, we created an interactive web application that allows researchers and developers to filter our extensive pattern set depending on their needs. Key Finding 4: Choosing the right communication pattern does not follow a simple one-size-fits-all approach; in contrast, which communication is best depends on the specific requirements of a situation.
不同的个人评分也反映了某些交流模式的差异和两极分化性质。虽然隐私、信任和用户总体偏好的测量结果似乎大多相关（见图 2），但其他测量结果似乎并不遵循类似的模式：虽然一种交流模式可能会传达较高的隐私和信任度，但也可能被认为是令人不安或几乎不引人注意的。此外，高差异也说明了人们对某些模式的感知普遍具有很强的主观性。由于我们认识到不同测量之间存在差异，而且单个测量的重要性取决于具体情况的特征，因此我们创建了一个交互式网络应用程序，允许研究人员和开发人员根据自己的需求过滤我们的大量模式集。主要发现 4：选择正确的交流模式并不是简单的 "一刀切"；相反，哪种交流模式最好取决于具体情况的具体要求。

We used an online survey to understand users' privacy concerns towards domestic robots with increasing capabilities. While online surveys are an established method to elicit privacy concerns [31,50], and sometimes the only viable option when investigating future scenarios, they might suffer from biases caused by participants having to immerse themselves in the described future or participants indicating answers that might not reflect their actual behavior [25]. In real life, participants might be more considerate of the convenience provided by the robot, making them willing to trade some of their privacy for an increased quality of life [17]. Moreover, the generally high privacy concerns might also be attributed to participants' low familiarity with such robots. Indeed, prior work already showed that higher familiarity is linked to decreased privacy concerns . Consequently, it will be interesting to repeat our survey in the future to see how concerns shift as users become familiar with domestic robots.
我们利用在线调查了解用户对功能不断增强的家用机器人的隐私问题。虽然在线调查是诱发隐私顾虑的一种成熟方法[31,50]，有时也是调查未来场景时唯一可行的选择，但由于参与者必须沉浸在所描述的未来中，或者参与者给出的答案可能并不反映他们的实际行为[25]，因此可能会出现偏差。在现实生活中，参与者可能会更多地考虑机器人提供的便利，从而愿意以部分隐私换取生活质量的提高 [17]。此外，参与者对隐私的关注度普遍较高也可能是因为他们对这类机器人的熟悉程度较低。事实上，先前的研究已经表明，较高的熟悉度与较低的隐私顾虑有关 。因此，我们有必要在未来重复我们的调查，看看随着用户对家用机器人越来越熟悉，他们的担忧会发生怎样的变化。

For this investigation, we did not consider the technical feasibility or how easy the gestures are to implement; we only focused on the users' perspective and which patterns provoke the highest levels of trust. Yet, in practice, technical feasibility is an important factor to consider when deciding which communication pattern to adopt. Hence, we recommend that future work employs a more technical focus and discusses the feasibility of our retrieved patterns from this perspective.
在这项调查中，我们没有考虑技术可行性或手势的易用性；我们只关注用户的观点，以及哪种模式能激起最高程度的信任。然而，在实践中，在决定采用哪种交流模式时，技术可行性是一个重要的考虑因素。因此，我们建议今后的工作应更多地关注技术问题，并从这一角度讨论我们检索到的模式的可行性。

We limited our elicitation of communication patterns to the three most privacy-concerning capabilities. We argue that limiting our investigation was important to be able to conduct the studies. Moreover, offering interventions and communicating the state of the most privacy-relevant capabilities is an approach frequently followed by manufacturers - many smart device manufacturers only provide mechanisms to physically block the cameras or integrate hardware buttons to deactivate the microphone. Yet, in reality, smart home appliances, and especially future domestic robots, will have way more privacy-relevant sensors, and which sensors are perceived as privacy-relevant might differ by user. Thus, it will be interesting to investigate which of our patterns apply to a broader range of sensors and where we need new mechanisms. Moreover, as previously discussed, concerns go beyond the pure collection of data as outlined in Solove's [47] taxonomy of privacy harms. Hence, future investigations are needed following this taxonomy as prior research already did for less capable smart assistants, c.f. [1, 2].
我们对交流模式的调查仅限于三种最涉及隐私的能力。我们认为，限制调查范围对于开展研究非常重要。此外，提供干预措施并通报与隐私最相关的功能状态是制造商经常采用的方法--许多智能设备制造商只提供物理阻挡摄像头或集成硬件按钮来关闭麦克风的机制。然而，在现实中，智能家电，尤其是未来的家用机器人，将拥有更多与隐私相关的传感器，而且用户对哪些传感器被视为与隐私相关可能也不尽相同。因此，研究我们的哪些模式适用于更广泛的传感器，以及我们在哪些方面需要新的机制，将是非常有意义的。此外，正如前面所讨论的那样，我们所关注的问题超出了索洛夫（Solove）[47] 的隐私危害分类法所概述的纯粹的数据收集。因此，未来的调查需要遵循这一分类法，正如之前的研究已经针对功能较弱的智能助手所做的那样，例如 [1, 2]。

We showed the focus group participants examples, i.e., a mute button and a physical camera shutter, to clarify what we mean by communication patterns. While our results show that our participants came up with a wide range of diverse patterns, we still want to acknowledge that these examples might have introduced unintentional biases as we can not exclude that other examples, such as LEDs [39] or dialogues with the user [57], might have led to different or more diverse communication patterns.
我们向焦点小组的参与者展示了一些例子，如静音按钮和物理相机快门，以澄清我们所说的交流模式的含义。虽然我们的结果表明，参与者提出了多种多样的交流模式，但我们仍然要承认，这些例子可能会带来无意的偏差，因为我们不能排除其他例子（如 LED [39]或与用户的对话 [57]）可能会导致不同或更多样化的交流模式。

Lastly, we used an online survey to describe the communication patterns in Study III. While we are certain that this is a good approach to get a first impression of the feasibility of the gestures, and online surveys are also a typical method used to gather human's perception towards robots [29], how the patterns are actually perceived in real life might be different. Hence, it would be desirable to test a selection of the patterns using prototypes, for example, in a lab study setting.
最后，我们在研究三中使用了在线调查来描述交流模式。虽然我们确信这是获得手势可行性初步印象的好方法，而且在线调查也是收集人类对机器人感知的典型方法[29]，但在现实生活中，人们对这些模式的实际感知可能会有所不同。因此，最好使用原型（例如，在实验室研究环境中）对部分模式进行测试。

We conducted three studies: An online survey ( ), a focus group study ( ), and a final large-scale online survey to understand users' privacy concerns towards future domestic robots and develop communication patterns to intervene with and signify their sensor state. Through this, we found that (1) the more interaction and movement capabilities a domestic robot has, the more concerns it evokes; (2) these novel capabilities also enable completely new communication patterns; and (3) most of these diverse patterns score equally well across all measurements, meaning that pattern use depends on the situation. To help researchers and developers navigate our extensive set of communication patterns along the mentioned characteristics, we developed an interactive web app. Finally, we discuss our key insights for choosing the right communication pattern: (1) selecting the mechanism based on the robot's capabilities, (2) choosing familiar patterns whenever possible to foster understandability and trust, and (3) being wary of the potential pitfalls when using humanoid metaphors.
我们进行了三项研究：通过在线调查（ ）、焦点小组研究（ ）和最后的大规模在线调查（ ），我们了解了用户对未来家用机器人隐私的担忧，并开发了干预和表示其传感器状态的交流模式。通过这些研究，我们发现：(1) 家用机器人的交互和移动能力越强，它所引起的关注也就越多；(2) 这些新功能也使全新的交流模式成为可能；(3) 这些多样化模式中的大多数在所有测量中得分相同，这意味着模式的使用取决于具体情况。为了帮助研究人员和开发人员根据上述特征浏览我们大量的交流模式，我们开发了一个交互式网络应用程序。最后，我们讨论了选择正确交流模式的关键见解：(1) 根据机器人的能力选择机制；(2) 尽可能选择熟悉的模式，以促进可理解性和信任感；(3) 在使用仿人隐喻时要警惕潜在的陷阱。