DAD-FL: Defending Federated Learning Against Poisoning Attacks via Cosine Similarity and Gaussian Mixture Modeling

Abstract

Federated learning (FL) enables collaborative model training without centralizing raw data, but its open participation mechanism makes the training process vulnerable to poisoning attacks. In realistic deployments, this problem is further aggravated by non-independent and identically distributed (Non-IID) data, because benign clients may naturally produce diverse updates that resemble malicious behavior. As a result, many existing defenses either over-filter benign clients or fail to detect attacks disguised as data heterogeneity. We propose DAD-FL, a dynamic defense framework for poisoning-resilient federated learning. DAD-FL first measures cosine similarity among client updates to identify suspicious directional inconsistencies, then projects candidate updates into a low-dimensional feature space and applies a Gaussian mixture model (GMM) for probabilistic anomaly detection, and finally performs robust aggregation over trusted clients only. By combining direction-aware screening with distribution-aware modeling, the proposed framework adapts to changing update patterns under Non-IID settings. Experiments on Fashion-MNIST show that DAD-FL improves attack detection while maintaining global model utility under different malicious-client ratios.