Abstract
While Simplicial Neural Networks (SNNs) are powerful tools for capturing complex, higher-order relationships in data, they come with a significant privacy trade-off. This research demonstrates that SNNs are more vulnerable to Graph Reconstruction Attacks (GRAs) than standard Graph Convolutional Networks (GCNs) or Graph Attention Networks (GATs). Our findings show that attackers can reconstruct original graph structures from SNN representations with over 98% accuracy, highlighting a critical need for privacy-preserving mechanisms in higher-order graph learning.
Introduction
Graph Neural Networks (GNNs) have become standard for analyzing social networks, biological data, and recommendation systems. These graphs often contain sensitive information, such as private friend lists or confidential professional connections. When a user's device computes a node representation (embedding) and transmits it to a server, there is an implicit assumption that the raw data remains hidden.
This paper investigates edge privacy: the risk that an adversary could use these output representations to reverse-engineer the original graph structure (the adjacency matrix). Specifically, we examine whether newer, more complex models like Simplicial Neural Networks—which encode interactions between groups of nodes (simplices) rather than just simple pairs—are more or less "leaky" than traditional pairwise models.
Methods
We evaluate privacy leakage by proposing a novel Graph Reconstruction Attack (GRA) using a "decoder-only" approach. In this setup, an adversary gains access to the node representations and a small portion of the original graph (non-confidential edges). The attacker then trains a graph decoder to minimize the difference between the reconstructed graph and the known partial graph.
We compared the vulnerability of three types of representations:
- Graph Convolutional Networks (GCN): Standard models encoding pairwise relationships.
- Graph Attention Networks (GAT): Models that weight the importance of neighboring nodes.
- Simplicial Neural Networks (SNN): Models using a combinatorial Laplacian to encode higher-order interactions (simplices), capturing the "shape" of the data.
The study utilized five standard real-world graph benchmarks: CiteSeer, Cora, PubMed, Amazon Computers, and Amazon Photo.
Key Results
The success of the reconstruction attack was measured using the Area Under the Curve (AUC) score. A higher score indicates a more successful recovery of hidden edges.
As shown in the experimental results, SNNs were consistently the most vulnerable across all datasets. On the CiteSeer dataset, the GRA achieved a staggering 98.33% AUC when attacking SNN outputs. Furthermore, the proposed GRA against SNNs consistently outperformed "Attack-6," a state-of-the-art baseline from 2021.
In contrast, GCNs demonstrated the highest resilience, with reconstruction AUCs typically ranging between 88% and 91%, significantly lower than the SNN results. GATs fell between GCNs and SNNs in terms of vulnerability.
Discussion and Conclusion
Our findings reveal a Complexity Paradox: the very feature that makes SNNs descriptive—their ability to capture multi-node interactions—is what makes them a privacy liability. Because SNN representations contain richer structural information, they provide a clearer "blueprint" for an attacker to follow. This study proves that "representation inversion" is a viable threat; simply sharing embeddings is not a sufficient privacy guarantee if the model is sophisticated.
Limitations and Future Work
This study focused on 1-simplices ($d=1$). Future research should explore if higher-dimensional simplices ($d \geq 2$) further increase leakage. Additionally, there is an urgent need for defense mechanisms, such as Differential Privacy or specialized regularizers, to enable the use of higher-order node information without sacrificing the security of the underlying graph structure.
Summary updated