Privacy-Preserving In-Vehicle Intrusion Detection via Nearest Centroid Classification Under Homomorphic Encryption

Intrusion Detection Systems are an essential component of modern vehicles, yet the data they process may expose sensitive information about drivers as well as proprietary information from the vehicle manufacturer. In this context, we explore two privacy-preserving solutions for intrusion detection on in-vehicle networks. The first is based on an existing decision tree proposal, while the second is a lightweight approach that we implement to reduce computational demands by performing nearest centroid classification directly on the encrypted data, serving as a baseline for evaluating the performance of privacy-preserving classification. Both approaches rely on the CKKS fully homomorphic encryption scheme, which is based on the Ring Learning With Errors (RLWE) problem and provides post-quantum security. This ensures that only encrypted data is hosted and processed by the cloud provider. As expected, these privacy guarantees come at a substantially higher computational cost than their unencrypted counterparts. Therefore, we compare these approaches with their traditional counterparts both in terms of intrusion detection performance and computational overheads. The evaluations are conducted on three existing datasets for vehicle intrusion detection, which contain traffic from real vehicles, providing practical insights on the current limitations of privacy-preserving intrusion detection on vehicular data with fully homomorphic encryption.