Evaluation of Label-Flipping Attacks on Federated Anomaly Detection in xG Networks

Abstract

Federated learning (FL)-based anomaly detection in next-generation (xG) networks is highly vulnerable to data poisoning, yet the impact of poisoning on model behaviour and decision structure remains poorly characterised. In this work, we present a system-level evaluation of label-flipping attacks across 306 configurations, spanning two datasets, multiple federation sizes, and complementary attack strategies. While prior studies typically rely on scalar metrics, we jointly analyse accuracy and Attack Success Rate (ASR), and introduce a scorespace analysis to characterise how poisoning alters the internal structure of anomaly predictions. This reveals three structurally distinct poisoning phases: stable, boundary degradation, and inversion. We observe a failure mode not previously reported in the literature: under strong poisoning, the model inverts, misclassifying attacks as benign while aggregate accuracy may remain deceptively acceptable during the early stages of failure. We further show that ASR rises measurably before accuracy degrades, providing an early-warning signal invisible to conventional evaluation. Together, these findings demonstrate that scalar metrics underestimate adversarial impact in FL, and establish distribution-level analysis as a complementary tool for monitoring deployed federated anomaly detection systems.