Multi-Node Attack and Defense Toolkit for Federated Learning : Exploration of Adaptive Aggregation Strategies

Sammanfattning

Federated learning (FL) collaboratively trains models without centralizing raw data, however its openness for clients invites poisoning attacks that can degrade or hijack the global models. This thesis develops a Kubernetes-orchestrated, multi-node attack toolkit for FL that couples the production-grade FEDn framework with a library of pluggable adversaries and defenses. The platform scales from a handful to hundreds of containerized clients via a one-line change and supports realistic features such as Dirichlet non-IID data splits, late client joins, various adversary activation and finally mitigation strategies. Using this test-bed, 180 controlled experiments were executed that crossed two attack families (static label-flipping and Little-Is-Enough), four data-partition regimes, two participation schedules, and three attacker densities (10%, 20%, 30%). In addition to a FedAvg baseline, this thesis benchmarks three state-of-the-art robust aggregation rules: Multi-KRUM, Trimmed-Mean, and Divide-and-Conquer. Furthermore, introduces EE-Trimmed-Mean, an adaptive rule that marries coordinate trimming with an ε-greedy trust score. The results confirm that no single static rule suffices across all conditions: Multi-KRUM can at times struggle to learn, whereas Trimmed-Mean does never fully exclude a malicious client update. EE-Trimmed-Mean achieved relatively good performance, while it stroke a balance of both including late benign client updates as well as filtering out malicious client updates. Adaptive, aggregation is a promising path towards resilient FL, it preserves honest diversity while achieving high malicious exclusion, whilst not relying on hard static thresholds. The open-source toolkit released with this thesis lowers the barrier for rigorous, reproducible security research in federated learning and provides a launch-pad for future work in the field of improving robustness and privacy in FL.