Collaborative Privacy: Using Social Trust and Bandit Algorithms to Protect Co-owned Data

TL;DR

Sharing a photo on Facebook often involves more than just the person who hits "post." This paper tackles the "co-owned data" problem by proposing a trust-weighted voting system. By treating the decision to post as a Multi-Armed Bandit problem, the authors show that users can maximize their social sharing benefits while minimizing privacy leaks through a dynamic, reputation-based trust model.

The Problem: The Unilateral Posting Trap

In current Online Social Networks (OSNs), the "owner" of a data item (the uploader) usually has sole control over its visibility. If User A posts a photo of User B, User A’s settings prevail. If User A shares it with colleagues while User B considers it private, User B suffers a privacy loss.

Prior attempts to solve this usually required a "trusted mediator" (like the OSN provider) to resolve conflicts. However, users are often reluctant to share their full privacy preferences with a central entity, and static rules fail to capture the nuances of evolving social trust.

Methodology: Trust as an Incentive Layer

The authors introduce a decentralized mechanism where the owner solicits "votes" from stakeholders before posting.

1. Trust-Weighted Voting

Instead of a simple majority, the "aggregated opinion" ($\bar{b}$) is calculated by weighting stakeholder votes by the owner's trust in them ($t_{os}$): $\bar{b}=\frac{{\sum }_{s\in S_d}{b}_s{t}_{os}}{{\sum }_{s\in S_d}{t}_{os}}$ The data is only posted if $\bar{b}$ exceeds a threshold $b_{th}$.

2. Dynamic Trust Evolution

Trust isn't static. It acts as a "social currency":

• Loss: If an owner ignores a stakeholder's privacy (causing a leak), the stakeholder’s trust in the owner drops.
• Gain: If the owner respects the stakeholder's preferences, trust increases.

This creates a feedback loop: if you violate my privacy today, I will trust you less tomorrow, and consequently, I will ignore your privacy preferences when I post co-owned data in the future.

The Balancing Act: The Bandit Approach

A user faces a constant trade-off: Data Sharing Benefit vs. Privacy Loss.

• Setting $b_{th}$ too low leads to constant sharing but massive reputation loss and future privacy risk.
• Setting $b_{th}$ too high preserves privacy but stifles the social utility of the network.

The authors formulate the selection of $b_{th}$ as a Multi-Armed Bandit (MAB) problem. Specifically, they use the Upper Confidence Bound (UCB) algorithm. The user "explores" different thresholds (arms) and "exploits" the one that delivers the highest payoff (Benefit - Privacy Loss).

Experiments: Trust Wins

The researchers tested their model on synthetic scale-free and small-world networks, as well as real-world Facebook data.

Key Findings:

• Altruism is Selfishly Beneficial: Users who were more considerate of others (solicited opinions) ended up with lower cumulative privacy loss because they maintained a higher reputation.
• Trust Outperforms Equality: Weighting votes by trust resulted in better privacy outcomes than simple majority voting.
• UCB Adaptability: As shown in the performance graphs, the UCB policy consistently outperformed fixed or random strategies, effectively learning the "sweet spot" for sharing.

Deep Insight & Conclusion

The core takeaway of this work is the mathematical validation of "Social Reputation" as a mechanism for privacy. By linking the act of posting to a dynamic trust value, the authors transform privacy management from a static permission check into a strategic, long-term game.

The use of Bandit Algorithms is particularly clever here because it acknowledges that "perfect privacy" is not the goal of a social network—"optimized sharing" is. While the model currently focuses on binary "post/don't post" decisions, the framework could easily scale to more granular access control levels in the future.