Support vector machines (SVMs) are widely used for online inference in AI enabled cyber-physical systems (CPS) because of their strong generalization ability and effectiveness on high-dimensional, nonlinear data. To reduce deployment costs on edge and end devices and improve service availability, SVM inference is often outsourced to the cloud. However, if the cloud server is semi-honest or malicious, the messages exchanged during inference may leak sensitive information about both user inputs and model parameters. In addition, a malicious server may return incorrect or forged inference results, which can mislead downstream decisions and threaten system security. It is therefore important to develop SVM inference mechanisms for untrusted cloud environments that provide privacy, verifiability, and efficiency. To this end, we propose two privacy-preserving inference schemes for adversaries with different capabilities. For the semi-honest setting, we first propose OPVSVM, an efficient privacy-preserving inference scheme based on secret sharing. OPVSVM protects user data and intermediate values and reduces computation and communication overhead by optimizing key inference operators. For malicious cloud servers, we further propose WPVSVM, a verifiable privacy-preserving inference scheme that sup-ports result verification without revealing sensitive information. Experimental results show that, compared with representative baselines, both OPVSVM and WPVSVM improve runtime efficiency and demonstrate practical performance for secure cloud-based SVM inference in AI-enabled CPS.

privacy-preserving, support vector machine, cloud computing, cyber-physical systems, IoT devices