AdaptKrum: privacy-preserving federated intrusion detection for industrial IoT under differential privacy noise and poisoning

Abstract

Industrial Internet of Things (IIoT) intrusion detection relies on traffic features that may still expose production schedules, device fingerprints, and process behavior after payload removal. Centralized training remains an accuracy upper bound, but it is difficult to operate when packet captures cannot leave the local network. AdaptKrum is a sample-level DPprotected federated intrusion-detection framework in which gateways keep raw traffic locally and upload only perturbed, compressed model deltas. The framework combines DP-SGD, Top-k plus 8-bit gradient compression, and Krum-style distance filtering with trimmed-mean aggregation. Its main design issue is practical: Gaussian DP noise pushes benign gradients farther apart, causing fixed-threshold Krum-style filtering to reject normal gateways. AdaptKrum estimates the effective accumulated DP-noise variance after local training, applies an engineering correction to retained sparsecoordinate distances, and rejects clients only when the corrected score exceeds an adaptive threshold. A 12-week physical deployment from March 1 to May 24, 2024 used 20 Jetson Nano B01 gateways across four workshop areas. At epsilon = 4, AdaptKrum achieved macro-F1 = 0.909, within 5.2 percentage points of centralized training, while reducing fullnetwork upload-only traffic from 6.40 to 1.88 MB per round. Under 30% model poisoning, it sustained macro-F1 = 0.851 and reduced benign-client false positives from 0.500 with fixed-threshold Krum-style filtering to 0.143.