Janus: Dual-Sided Communication and Privacy Optimization for Federated Learning in Autonomous Vehicles

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Federated Learning (FL) enables collaborative model training across autonomous vehicles (AVs) without sharing raw driving data, but excessive uplink/downlink communication over bandwidth-constrained vehicular-to-everything links and privacy leakage through gradient inversion attacks limit its deployment. This paper proposes Janus, a unified FL framework combining adaptive Top-K gradient compression with error feedback on the uplink, genuine additive homomorphic encryption (HE, not a noise-based proxy) for secure aggregation, and compressed global-model broadcasting on the downlink. On real KITTI object-detection labels with 20 non-IID clients, Janus matches FedAvg/FedProx/QSGD accuracy (95.34% ± 0.01% vs. 95.35% ± 0.03%, 3 seeds) while cutting combined communication by 61.8%; the same saving (61.0%) is independently reproduced on CIFAR-10. A real Paillier HE implementation is verified exact (error below 10⁻⁶), and a genuine gradient-inversion attack (Deep Leakage from Gradients) shows raw gradients are reconstructed almost perfectly (relative error 0.0026) versus 0.0775 for Janus's compressed updates, with HE removing the attack surface entirely. The same communication saving is further reproduced on real KITTI camera images with a CNN roughly 24x larger than the tabular model and on a multimodal model fusing real camera and LiDAR features, alongside additional baselines (FedPAQ, DP-FedAvg), membership-inference and model-poisoning attacks, Markov-chain mobility, simulated network latency (61.9% reduction), an energy proxy (45.1% reduction), and a formal convergence analysis. To the authors' knowledge, this is the first vehicular-FL evaluation combining joint uplink-downlink compression with genuine HE-based privacy, validated through real cryptography, real privacy attacks, statistical testing, cross-dataset and cross-modality generalisation, and a convergence guarantee.