SecureFL-IoV: Differentially Private Federated Intrusion Detection for IoV Using a Two-Stage RF

Abstract

The Internet of Vehicles (IoV) faces a fundamental security and privacy challenge, as centralized intrusion detection requires sensitive Controller Area Network (CAN) traffic to be transmitted beyond the vehicle. Federated Learning (FL) keeps raw traffic local, but shared updates remain data-dependent and can leak information. Formal Differential Privacy (DP) addresses this gap. This paper introduces SecureFL-IoV, an FL architecture combining a two-stage Random Forest (RF) cascade with Gaussian output perturbation DP, validated on simulated clients and a cloud-hosted five-client AWS EC2 testbed. The cascade uses a binary-then-multiclass design where Stage 1 produces thresholded binary attack/benign meta-features that augment the input to a Stage 2 six-class classifier. On deduplicated CICIoV2024 signatures, this design achieves a centralized Macro F1-score of 79.99% ± 4.58% under nested cross-validation, and 70.97% on the held-out protocol used by the federated phases, improving over the 64% single-stage RF benchmark. In fivesite simulated federation with NVIDIA FLARE and XGBoost RF mode, SecureFL-IoV achieves a Macro F1-score of 68.74% without DP and 61.33% at $\varepsilon=20$ with Gaussian output perturbation. In the cloud-hosted Non-IID EC2 deployment at $\varepsilon=20$, it achieves a 55.48% six-class Macro F1-score and 60.60% over five detectable classes. The observed $\pm 16 \%$ runto-run variance shows that benchmark scarcity and Non-IID partitioning dominate real-world performance. We also provide a custom NVFlare aggregator fixing two correctness bugs.