Adaptive Fidelity-driven Reconstruction (AFR): a realistic threat model for spectral embedding leakage

ICLR 2026 Conference Submission14162 Authors

Keywords: Federated Graph Learning, Privacy Leakage, Spectral Embeddings, Graph Reconstruction, Robust Alignment, Local Graph Benchmark

TL;DR: AFR shows that noisy, fragmented spectral embeddings in Federated Graph Learning can still be exploited to recover large, high-fidelity graph structures, elevating spectral leakage from a theoretical concern to a practical privacy threat.

Abstract: The exchange of structural representations in Federated Graph Learning (FGL) creates a potent channel for privacy leakage. While theoretical graph reconstruction is possible, existing attack models are brittle, as they hinge on an unrealistic assumption: perfect, noise-free local data. This paper elevates that theoretical threat to a practical reality. We introduce AFR (Adaptive Fidelity-driven Reconstruction), a robust new attack model that abandons idealized assumptions. Instead of assuming data quality, AFR actively measures and exploits it. The algorithm first quantifies the reliability of each local patch via a novel fidelity score, combining a spectral signal-to-noise ratio with structural entropy. This score then guides a robust assembly process that uses RANSAC-Procrustes to tolerate outliers and adaptive stitching criteria to manage uncertainty. Instead of a single, perfect graph, AFR recovers large, high-fidelity, and internally consistent islands from the most trustworthy data. Experiments on the LoGraB benchmark show that AFR successfully reconstructs significant topology in challenging, noisy regimes where idealized models fail completely. Our work thus promotes spectral leakage from a theoretical possibility to a practical and potent threat. Our source code is anonymously available at: https://anonymous.4open.science/r/AFR-ICLR-submission.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 14162