Federated Learning Security: An Incremental Distributed Attack Approach

Federated Learning (FL) is esteemed for its efficiency and privacy preservation in IoT applications, yet its popularity raises concerns about susceptibility to adversarial attacks. We propose a novel federated and incremental attack initiated by the aggregation server, targeting neural network training on the client’s side. This agnostic attack focuses on intermediate data structures used in FL, tested in a secure cognitive city application for biometric authentication. Two measures of attack stealthiness are proposed: the difference in training time and the average of evaluation loss change measure. Experimental results show attacked networks with higher training loss change and an effectiveness measure of $62 \%$. Despite its efficacy, the attack maintains stealthiness with minimal impact on federated training time and evaluation loss, affirming its effectiveness. The code is publicly accessible at: https://github.com/Artifitialleap-MBZUAI/Incremental Federated Attack on Neural Network Training.