Privacy-Preserving K-Means Clustering for Vehicular Driving Behavior Analysis

The rapid development of intelligent transportation systems has led to the generation of massive vehicular data from onboard sensors, GPS devices, and driving logs, paving the way for advanced driver behavior profiling. To extract insights from such data, clustering techniques, particularly k-means, are widely employed. However, traditional centralized k-means clustering for driving behavior analysis poses potential privacy leakage, as raw features are transmitted to the cloud server in plaintext. Existing privacy-preserving k-means schemes either rely on interactive homomorphic encryption protocols, which incur significant communication overhead, or adopt differential privacy, which reduces accuracy due to noise injection. To address these challenges, we propose PPKC, a non-interactive privacy-preserving k-means clustering framework tailored for vehicular driving behavior analysis. PPKC leverages an improved Paillier cryptosystem with a three-modulus construction, enabling secure Euclidean distance computation and cluster center updates directly over ciphertexts without decryption or repeated interaction with data owners. Each vehicle only uploads a single round of encrypted feature vectors, after which the cloud server can independently execute the complete k-means clustering process. We rigorously analyze the security of PPKC, proving its semantic security and indistinguishability under the IND-CCA assumption with unforgeable digital signatures. Extensive experiments on the highD dataset demonstrate that PPKC achieves high clustering accuracy with significantly lower communication and computation overhead than existing protocols.