Your Privacy My Cloak: Backdoor Attacks on Differentially Private Federated Learning
Prior research assumed differential privacy (DP) protects federated learning from backdoor attacks. This paper challenges that by introducing RING, an attack that exploits DP noise to hide malicious updates. RING crafts adversarial perturbations that look like noisy, benign updates while ensuring they cancel out during aggregation, allowing the backdoor signal to recover without detection.