FedGuard-IDS: A Byzantine-Resilient Federated Network with Secure Aggregation for Collaborative Intrusion Detection

Abstract

Network intrusion detectors trained inside one organisation fail on the traffic of another, yet organisations cannot pool their flows. A federation that shares only model updates is open to poisoning by a few dishonest participants, and the secure aggregation that hides each update also hides the evidence that robust aggregation needs. We present FedGuard-IDS, a federated network for collaborative intrusion detection that combines pairwise-masking secure aggregation over small random groups of clients with a server-side acceptance test on a small clean validation set. The server sees only masked uploads and group means; it clips the means, accepts only those whose updated model does not raise the loss on the clean set, averages the middle of the accepted means, and falls back on its own update when none passes. We evaluated it on two public NetFlow benchmarks, UNSW-NB15-v2 and ToN-IoT-v2, treated as two organisations with 10 clients each, against four attacks at 10, 20 and 30% malicious clients, with five seeds, fifteen aggregation baselines and server-only and centralised references (1,928 runs). Without attack, FedGuard-IDS reached an F1 score of 0.982, against 0.987 for a centralised network. In all 60 attacked runs its F1 was at least 0.971. FedAvg fell to 0.223 under sign flipping, and 13 further baselines fell below 0.8 in at least one setting, including bucketing in pairs followed by standard robust rules, which isolates the acceptance test as the protective component; FLTrust stayed near 0.96 but scored 0.59 to 0.77 on the smaller organisation. Attackers that know the gate and target the smaller organisation kept the overall F1 at 0.98 but lowered the recall of its attacks from 0.967 to about 0.92, and hurt other defences far more. Pairs of clients gave the best robustness, and a server's best guess of one client's organisation fell from 1.00 to 0.74. The evidence is limited to laboratory benchmarks and simulated secure aggregation.