Poisoning-Resilient Homomorphic Secure Aggregation for High-Dimensional Cybersecurity Gradients

Abstract

Federated Learning (FL) allows for collaborative model training across several geographically dispersed clients while protecting client data from other users. Although FL is susceptible to poisoning attacks, collusion, and Byzantine adversaries, which can result in decreased performance of the overall model, we present a lightweight FL framework that utilizes homomorphic encryption to protect against poisoning attacks while also providing client data protection and low computational overhead by integrating Additive Homomorphic Encryption (AHE), Trust-Weighted Aggregation (TWA), Anomaly Scoring, and Gradient Clipping. The proposed framework has been extensively simulated with the MNIST and CIFAR-10 datasets. These simulations have demonstrated that the proposed framework can maintain greater than 95% of the overall model's accuracy when the number of malicious clients is as high as 40%. Local epochs are processed at the client side in 15-20 milliseconds; server-side secure aggregation was found to be two times slower than plaintext aggregation; and each round of communication averages approximately 4 MB/client of data transferred. The simulation results provide evidence that the proposed framework will provide robust convergence, scalability, and resilience to a variety of types of adversary behaviors. Existing approaches address either privacy or robustness, but this work proposes a unified framework that simultaneously ensures privacy (via homomorphic encryption) and robustness (via poisoning-resilient aggregation) for applications such as cybersecurity monitoring, IoT, and other privacy sensitive applications.