Mitigating Gradient Leakage in Asynchronous Federated Learning via Orthogonal Obfuscation

Abstract

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 Asynchronous FL (AsyncFL). This architectural discrepancy creates a critical security gap: the lack of coordinated updates allows adversaries to isolate single-client contributions, rendering standard privacy guarantees ineffective. In this work, we first quantify this exposure through a data reconstruction attack that exploits the temporal difference in global model states to recover sensitive client data with high fidelity. To mitigate these specific risks, we propose a defense-in-depth framework that combines hardware-enforced isolation with algorithmic obfuscation. Our solution utilizes Trusted Execution Environments (TEEs) to secure server-side aggregation and introduces a client-side mechanism: Gradient Obfuscation via Orthogonal Distribution (GOOD). This technique entangles private gradients with procedurally generated fractal distributions in the latent space, to mitigate gradient leakage during asynchronous FL aggregation. Extensive evaluations across multiple neural architectures demonstrate that our approach reduces the Structural Similarity Index (SSIM) of reconstructed data by 85% and increases reconstruction error by up to 3×, all while maintaining model utility within 1% of the non-private baseline.