Stealthy Backdoor Attack in Federated Learning via Adaptive Layer-wise Gradient Alignment

We study the tradeoff between the strength and detectability of backdoor attacks in federated learning. Layer-wise alignment uses the previous global update as a reference to constrain malicious updates adaptively, revealing vulnerabilities across eight robust defenses and a range of datasets and networks.

Recommended citation: Yang, Q., Yan, P., Wu, X., et al. (2025). "Stealthy Backdoor Attack in Federated Learning via Adaptive Layer-wise Gradient Alignment." ICCV 2025.
Download Paper