An Attention-Driven CNN-BIGRU Architecture For Privacy-Preserving Federated Intrusion Detection In Internet Of Vehicles Can-Bus Networks

Abstract

The growing number of Internet of Vehicles (IoV) ecosystems has greatly increased the options for cyberattack on Controller Area Network (CAN) buses. Centralized Intrusion Detection Systems (IDS) are incapable of performing effectively in a distributed vehicular environment because they rely on centralized data collection, have single points of failure and have significant privacy weaknesses. The framework proposed in this paper is a new privacy-preserving federated IDS framework that uses a hybrid CNN-BiGRU-Attention deep learning architecture and an Adaptive Weighted Input (AWI) aggregation function. This hybrid architecture includes the use of 1D convolutional layers to extract local spatial features from CAN payload bytes; two independent BiGRUs to capture forward and reverse temporal dependencies in sequences of network traffic; and a soft attention mechanism to dynamically weight the most discriminative temporal segments. The AWI aggregation replaces conventional FedAvg with continuous trust-based weighting combining cosine similarity and norm deviation analysis, enabling adaptive suppression of adversarial client contributions. Experimental evaluation on the CIC-IoV 2024 benchmark dataset (281,644 labeled CAN-bus samples with DoS, spoofing, and replay attacks) demonstrates 98% accuracy, precision, recall, and F1-score in the federated setting, with an ROC-AUC of 0.998. Five-fold cross-validation confirms stable generalization at 97.14% ± 0.29%. A paired t-test (p < 0.05) confirms superiority over CNN, LSTM, SVM, Random Forest, XGBoost, and FL-CNN baselines. The ablation study establishes that the dual-stage defense and AWI mechanism are the largest individual contributors under adversarial conditions.