Lossless Privacy for Federated Security: Homomorphically-Encrypted Federated Learning Across IoT Intrusion Detection, Keystroke Dynamics, and Risk-Based Authentication

Abstract

Federated Learning trains a shared model without centralizing raw data, yet the de-facto FedAvg protocol still reveals each client's plaintext model update to the aggregation server, opening the door to gradient-inversion and membership-inference attacks. We present HE-FedSec, a federated framework that performs secure aggregation directly on ciphertexts with the CKKS homomorphic-encryption scheme, so the server never observes any individual update. We evaluate HE-FedSec uniformly across three complementary, privacy-sensitive security domains under non-IID client partitions, which include 15-class network intrusion detection on Edge-IIoTset, 51-subject keystroke-dynamics identification on the CMU benchmark, and binary malicious-login detection on the large-scale RBA dataset. Relative to plaintext FedAvg, HE-FedSec is near-lossless on the two multiclass tasks, test accuracy changes by at most 0.26 percentage points, while on the imbalanced RBA task the ROC-AUC gap is only 0.11 points. This behavior is consistent with the small CKKS numerical error introduced into the weighted-mean aggregation. This privacy guarantee incurs a $21-27 \times$ communication expansion while introducing a minor cryptographic latency of only a few to tens of milliseconds per round. By contrast, an update-level differential-privacy baseline at a comparable privacy posture loses up to 5.9 accuracy points and converges more slowly. We release the complete, reproducible pipeline and benchmark.