Skip to content

Aegis: Generative Gradient Masking for Privacy-Preserving Medical Federated Learning

Sep 2026 · arXiv.org · Vol abs/2609.38339 · 0 citations · 49 references
Computer Science

TL;DR

Aegis is proposed, a principled client-side defense that breaks this dilemma without perturbing patient data or modifying the FL protocol, offering a practical privacy primitive for medical FL.

Abstract

Federated learning (FL) has become a foundational paradigm for multi-institutional medical AI, allowing hospitals and research centers to jointly train diagnostic models without exchanging patient records. This privacy promise, however, is increasingly contested: a malicious or honest-but-curious server can launch model inversion attacks (MIAs) that reconstruct private patient images directly from shared model updates, and recent scalable, closed-form attacks penetrate even secure aggregation at clinically realistic batch sizes. Existing defenses face an unsatisfactory dilemma. Gradient-perturbation methods such as differential privacy and pruning trade away the diagnostic accuracy on which clinical reliability depends, while cryptographic protocols add system complexity yet still leave updates exposed to these scalable attacks. We propose Aegis, a principled client-side defense that breaks this dilemma without perturbing patient data or modifying the FL protocol. Our key insight is that the success of every known MIA is fundamentally bounded by the local batch size relative to the model's leakage capacity; once this limit is exceeded, distinct samples collide and reconstructions collapse into indistinguishable mixtures. Aegis turns this universal bottleneck into a defense: each client superimposes onto its real update a masking gradient computed on locally synthesized, task-relevant data, deliberately pushing the effective batch beyond the attack's recovery capacity. We complement the design with theoretical convergence guarantees under standard convex assumptions and evaluate Aegis on MNIST, CIFAR-10, and three MedMNIST modalities (chest X-ray, abdominal CT, colon pathology). Aegis neutralizes three state-of-the-art MIAs while preserving model utility and incurring only modest overhead, offering a practical privacy primitive for medical FL.

View source

Similar papers

Open access Oct 2026

Efficient Privacy-Preserving Medical Cross-Silo Federated Learning.

This approach holistically rethinks the entire FL-SMPC pipeline, incorporating concurrency at critical stages to improve runtime efficiency while preserving strong security guarantees for the aggregated global model on the server side-an often overlooked aspect in existing SMPC approaches.

Benjamin Hamm, Yannick Kirchhoff, M. Rokuss et al. · 2 citations
Review Open access 2026

Privacy and Security in Knowledge Distillation for Federated Learning: A Survey

This survey examines KD-based FL through a privacy and security lens, and provides a precise framework for evaluating what KD-FL methods actually guarantee and what they only make harder for an adversary.

Hamza Reguieg, Essaid Sabir, M. El Kamili · 0 citations
Conference Aug 2026

Privacy-Preserving Federated Learning with Differential Privacy: Methods, Threats, and Future Directions

Federated Learning (FL) and Differential Privacy (DP) have become fundamental techniques for building privacy-preserving and decentralized machine learning systems. FL supports collaborative training of models over distributed devices and organizations without exchanging raw data and DP enhances privacy guarantees by a...

Prabushankar K. C., B. Buvaneswari, P. Rajamani et al. · 0 citations
Conference Sep 2026

Mitigating Gradient Leakage in Asynchronous Federated Learning via Orthogonal Obfuscation

The Federated Learning (FL) paradigm, while promising for decentralized training, remains susceptible to privacy breaches via gradient leakage. While numerous defenses have been established for synchronous systems, these mechanisms often rely on batched aggregation to mask individual contributions, a property absent in...

Shamim Yazdani, Noah Guzinski, George Torres et al. · 0 citations

Related blog posts

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.