Fed3D: Enhancing Security in Federated Learning with Dataset Distillation

Dataset Distillation (DD) compresses large datasets into compact representations while preserving performance, offering substantial benefits for Federated Learning (FL). However, using distilled datasets introduces new security vulnerabilities, as adversaries can easily embed backdoors into the distilled data. In this paper, we extend existing backdoor attack strategies in DD to the Federated Learning context (DD-FL) and empirically demonstrate their effectiveness. To address these threats, we propose Fed3D, the first defense algorithm specifically designed for DD-FL. Fed3D incorporates a dual-layer defense mechanism, combining intra-client diversity detection with inter-client clustering based on reconstructed feature representations. Comprehensive experiments show that Fed3D effectively reduces attack success rates (<1.5%) while maintaining the performance of distilled datasets (<1.1%). These results establish Fed3D as a robust and promising solution for mitigating backdoor attacks in DD-FL systems.