New Attack Bypasses Krum Aggregation in Federated Learning.

Robust aggregation methods, such as Krum and Multi-Krum, are widely used in federated learning to protect against adversarial client behavior. These methods select client updates that are geometrically closest to the majority, assuming benign updates form a compact cluster. However, this reliance on geometric properties can be exploited by sophisticated adversaries. This research introduces the Krum-Proxy attack, a novel backdoor injection strategy designed to consistently bypass Byzantine-robust aggregation. Unlike simpler scaling or constraining methods, Krum-Proxy actively optimizes malicious updates to blend into the dense core of the benign update distribution. It achieves this through a two-stage optimization process that separates task-specific attack objectives from geometry-aware refinement, utilizing a nearest-neighbor proxy, stochastic reference modeling, and anchor-guided alignment. To maintain stealth, the attack incorporates a projection mechanism that constrains adversarial updates within realistic norm and variance bounds. Experiments on standard federated learning benchmarks demonstrate that Krum-Proxy achieves higher attack success rates while preserving the clean accuracy of the model, thereby highlighting a significant vulnerability in distance-based aggregation rules against selection-aware adversaries.