PPRRP: Balancing Privacy and Utility in Weighted Social Networks via Rotation Perturbation

TL;DR

Social network data is a goldmine for research, but sharing weighted graphs (where edges represent interaction frequency or strength) poses severe privacy risks. This paper proposes PPRRP, an algorithm that uses rotation perturbation to mask edge weights. Unlike traditional methods that distort the network structure, PPRRP keeps the "feel" of the data (shortest paths, clustering) nearly identical to the original while ensuring the actual numbers are safely hidden.

Context & Motivation: The Weight of Privacy

In a social network, it’s not just about "who knows whom," but "how well they know each other." These weights are sensitive. However, if you simply add random noise to these weights, you break the shortest path properties—the very things researchers want to study.

The authors identify a critical gap: existing methods like Random Basis Transformation (RBT) introduce too much structural "noise." Their insight is that Rotation Perturbation can change individual values while preserving the relationships (inner products and distances) between vectors in a higher-dimensional space.

Methodology: The Geometry of Anonymity

The PPRRP algorithm operates through a multi-step process:

1. Vector Set Construction: The network is transformed into a matrix $M(G)$, where each node's relationships represent a vector.
2. Rotation Perturbation Matrix: A matrix is generated using a random function $r(i,j)=\beta imes{q}_{ij}$. This matrix rotates the original data vectors.
3. Privacy Control: The degree of perturbation is measured using the variance of the difference between original and perturbed datasets ($Var(X-X^{\prime })$).

The beauty of the rotation is that it is an isometric transformation. While the individual values $e_{ij}$ change drastically, the relative distances $dist(v_i,v_j)$ are largely preserved, which is why the network topology remains stable.

Experimental Validation: SOTA Performance

The authors tested PPRRP on two distinct datasets: Lesmis (small, dense) and Hepcollab (large, sparse).

1. Privacy Quality

The privacy strength was evaluated by looking at the variance and the minimum difference ($p_{max}$). The method ensures that even if an attacker knows some background information, the rotation makes it mathematically difficult to "reverse-engineer" the original weights.

2. Data Utility (The Real Winner)

The table below highlights how PPRRP keeps the network metrics much closer to the "Original" than its competitors:

As shown, the Average Shortest Path and Clustering Coefficient under PPRRP are nearly identical to the original data, proving that the perturbed graph is still highly useful for social science research.

Critical Analysis & Conclusion

Takeaway

PPRRP proves that you don't have to destroy data quality to protect privacy. By moving from "scalar noise" to "vector rotation," we can satisfy both the privacy officer and the data scientist.

Limitations & Future Work

The current approach assumes a static network. In real-world scenarios, social networks evolve. Future research should investigate how these rotation matrices perform when new edges are added over time—ensuring that the "rotation" doesn't become predictable across different snapshots of the same graph.

This work sets a strong baseline for the next generation of privacy-preserving graph publishing.