Taking Situation-Based Privacy Decisions: Privacy Assistants Working with Humans

Nadin Kökciyan, Pinar YolumIJCAI 2022 · DOI

TL;DR

This paper introduces a novel agent-based privacy assistant (PAS) that makes privacy decisions for users in the context of the Internet of Things (IoT). Unlike traditional approaches that rely on extensive user interaction or pre-defined rules, PAS dynamically identifies contexts from situations and assesses trustworthiness based on these contexts, even for previously unseen devices. This allows for more personalized and less burdensome privacy management in the increasingly complex IoT landscape.

What Problem Does This Paper Address?

The proliferation of Internet of Things (IoT) devices has created a significant challenge for users in managing their privacy. Traditional consent mechanisms, where users must agree to privacy policies for each device and service, impose an overwhelming cognitive load. This often leads users to ignore the details, rendering the consent process ineffective. Furthermore, the dynamic and interconnected nature of IoT means users frequently encounter new devices and situations for which they have no prior experience or established trust. Existing privacy assistants can learn user preferences, but the authors argue for a system that can make decisions on behalf of the user in novel situations by understanding the context. The core problem is how to enable a privacy assistant to make informed, situation-specific privacy decisions for users, especially when dealing with unfamiliar IoT devices and contexts, without requiring constant user intervention.

Methods

The paper proposes a Situation-Based Privacy Model for privacy assistants (PAS). The key innovations lie in:

  • Context Identification: Instead of relying on pre-defined contexts, PAS dynamically derives contexts by grouping similar privacy situations. A privacy situation is defined by features of the IoT device and the interaction (e.g., purpose, data type, location).
  • Contextual Trust Assessment: PAS assigns trust not to individual devices but to contexts. It uses subjective logic to compute opinions (degrees of belief, disbelief, uncertainty) about contexts based on the user's past positive and negative experiences within those contexts. This allows for trust assessment even without prior interaction with a specific device.
  • Multi-context Decision-Making: For unseen situations, PAS first categorizes them into multiple contexts with varying degrees of confidence. It then computes a weighted probability expectation value across these contexts to make a decision (allow or deny data collection).
  • Delegation Mechanism: PAS identifies ambiguous situations (high conflict between belief and disbelief in a dominant context) and delegates the decision back to the user, preventing over-reliance on potentially flawed automated decisions.
  • Personalization: The model allows for personalization by adjusting parameters (like conflict thresholds) based on user groups or individual behavior patterns.

Datasets

The authors used an anonymized IoT dataset collected through surveys from users. This dataset contains 380 privacy scenarios, each described textually and accompanied by user responses regarding data collection. The dataset was used for both training the context classification model and evaluating the PAS performance. No new benchmarks were introduced; the dataset was used as is.

Key Insights and Ideas

  • Context is King (for Privacy): The paper strongly suggests that privacy decisions are more effectively made by understanding the context of a situation rather than just the specific device or a broad category. For instance, being in a library doesn't inherently dictate privacy preferences; the purpose of data collection in that library does.
  • Trust in Context, Not Just Entities: The novel approach of building trust in contexts rather than individual devices is crucial for IoT, where users interact with many devices infrequently. This allows for generalization and informed decisions even for new devices.
  • Dynamic Context Derivation is Essential: Pre-defining contexts is rigid. The ability to dynamically group similar situations into contexts allows the system to adapt to evolving user behaviors and new types of IoT interactions.
  • Balancing Automation and User Control: The delegation mechanism is vital. Identifying and deferring ambiguous decisions to the user is key to maintaining user trust and achieving higher overall accuracy than purely automated systems in uncertain scenarios.
  • Personalization is Non-Negotiable: Different users have vastly different privacy needs and risk tolerances. The model's ability to adapt its decision-making parameters based on user groups or individual behavior is critical for practical adoption.

Key Results

  • The multi-label classifier trained to infer contexts achieved high performance, with macro-averaged precision, recall, and F1-scores of 0.91, 0.88, and 0.90, respectively, and an average accuracy of 0.86.
  • The experimental results (Table 1) showed that the PAS-M (multi-context, no human intervention) generally outperformed PAS-S (single context, no human intervention), especially for users with more conservative privacy stances (Group 3).
  • When PAS delegated decisions to the user (PAS with human intervention), accuracy consistently increased, highlighting the effectiveness of the conflict-based delegation strategy. For example, accuracy could increase to 0.89 in ambiguous cases.
  • The study demonstrated that different user groups (Behavioral Unconcerned, Pragmatists, Fundamentalists) required different settings for conflict thresholds (ψ) and confidence intervals (γ) for optimal performance, underscoring the need for personalization.

Limitations and Future Work

  • Approximation in Experience Bases: For the case study, the authors had to populate participants' experience bases using data from similar users due to the lack of extensive individual historical data. This is acknowledged as an approximation that could lower accuracy in real-world scenarios where only the user's own experiences would be available.
  • Limited Dataset Size: The evaluation used only 14 scenarios per participant, which is far less than a realistic setting.
  • Future Work:
    • Semantic Relations Between Contexts: Exploring how semantic relationships between contexts could inform decision-making order.
  • More Elaborate User Feedback: Incorporating richer feedback from users beyond simple "allow/deny" to gain deeper insights into their privacy concerns.
  • Incorporating Personal Values: Integrating personal values alongside contexts for more nuanced decision-making.

Summary updated

BibTeX
@inproceedings{kokciyan2022taking,
  author = {Nadin Kökciyan and Pinar Yolum},
  title = {{Taking Situation-Based Privacy Decisions: Privacy Assistants Working with Humans}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.24963/ijcai.2022/99},
}
Figure