Protecting Link Privacy for Large Correlated Social Networks: Beyond Independent Data Assumptions

TL;DR

In the landscape of social data, "no user is an island." Most privacy algorithms treat social links as independent events, but real-world networks are ripples of correlation. This paper introduces an improved perturbation-based approach that identifies "invisible" social ties using a tri-factor correlation model. By swapping edges based on these correlations, the authors achieve a more robust privacy shield without destroying the network’s statistical utility.

Problem & Motivation: The Myth of Independence

In graph data, privacy is often synonymous with "Edge Perturbation"—swapping link $A-B$ for $A-C$. However, if $B$ and $C$ are highly correlated (e.g., they share 15 mutual friends), the adversary can easily infer that $A$ is likely connected to $B$ even if the direct link is hidden.

The authors argue that previous works, such as those by Mittal et al., simply choose new neighbors from local structures. This fails because:

1. Structural Estimation: Sparse data can be used to estimate sensitive links.
2. Correlation Neglect: Indirect social ties (the "friends of friends" effect) are often ignored, leaving a trail for de-anonymization.

Methodology: Quantifying the Social "Gravity"

The core of the proposed method is a transition probability matrix that factors in three distinct social dimensions to define how "close" two people actually are:

1. Shortest Path Length ($D$): Based on "Six Degrees of Separation."
2. Common Neighbors ($E$): The more mutual friends, the higher the correlation.
3. Degree Punishment Factor: High-degree nodes (influencers) are treated as "weaker" individual ties because they connect to everyone, reducing the specific correlation between any two of their neighbors.

The Perturbation Workflow

For each vertex $u$, the algorithm:

• Deletes an existing edge $(u,v)$.
• Initiates a random walk from $v$ to a terminal vertex $z$, governed by the calculated transition probabilities.
• Adds a new edge $(u,z)$.

This process ensures that the "social distance" in the perturbed graph is balanced, making it difficult for an adversary to distinguish sensitive links from noise.

Experiments & Results: Winning in the Dense Crowd

The authors tested their approach on three datasets: CA-GrQc (Arxiv collaboration), CA-HepTh (High Energy Physics), and Wiki-Vote (Wikipedia voting).

Performance Gains

The most striking results occurred in dense graphs (Wiki-Vote). In sparse graphs, the lack of candidate paths makes perturbation difficult. However, in dense settings:

• Privacy Improvement: Lower "Privacy" scores (representing lower predictability for adversaries) were consistently achieved compared to standard random walks.
• Utility Preservation: Global properties like Average Degree and Degree Standard Deviation remained almost identical to the original graph, ensuring the data remains useful for research.

Table: Comparison of various parameter settings ($\alpha ,\beta ,\gamma$) on the Wiki-Vote dataset.

Critical Analysis & Conclusion

Takeaway

The paper successfully demonstrates that link privacy is a multi-dimensional problem. By effectively "balancing" the graph—reducing the variance of distances between nodes—the authors make the structural signal and the noise indistinguishable.

Limitations

• Sparsity Issues: As the authors admit, the algorithm struggles with very sparse graphs (random graphs) where few edges exist to provide alternative paths.
• Weight Sensitivity: The performance relies heavily on the weights $(\alpha ,\beta ,\gamma )$. While $\beta =1$ (focusing on common neighbors) worked best for dense graphs, there isn't a "one-size-fits-all" setting for every network type.

Future Outlook

As we move toward a more connected world, privacy methods must evolve to handle the correlation of everything. This work serves as a foundational step toward "correlation-aware" privacy, which could eventually be integrated into larger frameworks like Differential Privacy to provide formal mathematical guarantees alongside structural utility.