PQ-FL-IDS: A Post Quantum Safe Federated Edge Intrusion Detection Scheme underlying 6G

Abstract

Federated learning (FL) enables Intrusion Detection System (IDS) models to be trained across Sixth Generation (6G) edge nodes without transferring sensitive traffic records to a central server. However, the model updates exchanged during FL-based IDS (FL-IDS) training remain exposed to interception and harvest-now-decrypt-later attacks by quantum-capable adversaries. This paper presents PQ-FL-IDS, a post-quantum (PQ) FL-IDS scheme for 6G edge networks. It protects model-update exchange using a real, non-simulated Module-Lattice-Based Key-Encapsulation Mechanism (ML-KEM-512, NIST FIPS 203), the Advanced Encryption Standard in Galois/Counter Mode with a 256-bit key (AES-256-GCM), and per-round forward secrecy through ephemeral key pairs. Differentially Private Stochastic Gradient Descent (DP-SGD) with Rényi DP accounting further bounds record-level leakage from local updates. In each round, clients train locally with calibrated DP noise, encrypt the model update, and send only the protected payload for server-side aggregation. We evaluate PQ-FL-IDS on ToN-IoT with a 1D-CNN and FedAvg, and on CICIoT2023 with MobileNetV2-1D and FedProx, under baseline, PQC-only, and PQC+DP settings. ML-KEM-512 with AES-256-GCM adds less than 0.003% computational overhead relative to per-round training time and at most 3.5 KB of payload expansion, with no accuracy loss, and remains stable from 2 to 5 clients. The privacy analysis shows predictable ε growth with the subsampling rate and characterizes the privacy-utility frontier: formal budgets of ε = 0.9-2.4 are obtained for σ = 0.8-1.1, with a utility cost that depends strongly on task complexity.