Hear no evil: detecting gradient leakage attacks launched by malicious servers in federated learning

Researchers analyzed gradient leakage attacks launched by malicious servers in federated learning from a defense perspective, finding that such attacks struggle to simultaneously efficiently reconstruct clients' private data and remain stealthy enough to evade detection. In such attacks, the malicious server manipulates the global model to induce clients to upload gradient updates containing sensitive information; the authors call this the first comprehensive analysis of this class of attacks and their model manipulation techniques. This trade-off is particularly pronounced in realistic federated learning settings that include common normalization techniques and federated averaging, and the authors accordingly argue that although the attacks are theoretically concerning, they are inherently limited in practice and can often be detected through basic monitoring. The paper also proposes a simple, lightweight and broadly applicable client-side detection mechanism that flags suspicious global model updates before local training begins, providing a deployable defense for privacy-focused federated learning systems at minimal overhead.