FedCIPHER: Causal-Invariant and Privacy-Hardened Prototype Federation for Robust Network Intrusion Detection under Non-IID Traffic

Abstract

Collaborative network intrusion detection requires security operation centers, cloud platforms, and industrial edge domains to learn from distributed traffic without centralizing packet traces, asset identifiers, or attack evidence. Federated learning limits direct data movement, yet conventional parameter averaging is fragile when attack classes are rare, traffic distributions are non-independent and identically distributed (non-IID), and model updates can be inspected or poisoned. This paper proposes FedCIPHER, a causal-invariant and privacy-hardened prototype federation framework for multi-class network threat detection. Each client learns a gated causal representation by exchanging nuisance components between traffic samples and enforcing counterfactual prediction consistency. Instead of uploading full model gradients, the client releases clipped, orthogonally sketched first- and second-order class prototypes protected by distributed Gaussian noise and secure summation. An evidential uncertainty score and a class-conditional deviation statistic jointly control robust server aggregation, while a private global prototype bank regularizes locally personalized detectors. We provide a Rényi differential privacy bound, a bounded-influence analysis for malicious clients, and a convergence characterization that explicitly separates client drift, sketch distortion, and privacy noise. Experiments on UNSW-NB15, CICIDS2017, ToN-IoT, and BoT-IoT with 20 clients show that FedCIPHER obtains average Macro-F1 of 88.70% under Dirichlet concentration α = 0.3, exceeding FedAvg, FedProto, FEDIIR, and PROTEAN by 6.68, 3.69, 3.61, and 2.43 percentage points, respectively. At ε = 4 and δ = 10−5, membership-inference AUC is reduced to 0.566 while Macro-F1 decreases by only 0.42 points relative to the non-private variant. FedCIPHER also retains 82.83% Macro-F1 with 30% malicious clients and reduces per-round upload by 99.59% compared with full-model federation. These results indicate that causal prototype federation offers an effective balance among detection utility, privacy, poisoning resilience, and communication efficiency for cross-domain network data security.