Defending Against Gaussian Process Membership Inference Attack

As the deployment of AI systems in sensitive domains increases, they are becoming increasingly vulnerable to various cyberattacks. Among these, membership inference attacks (MIAs) pose a significant privacy threat by attempting to determine whether specific data points were included in a model's training set. In this work, we focus on Gaussian Process (GP) models, which are widely used in machine learning for their flexibility and ability to quantify uncertainty, and investigate their susceptibility to MIAs. To address this issue, we propose a novel defense mechanism based on random unitary transformation. Our method encrypts training and/or testing data using orthonormal random matrices, enabling secure GP regression without adding noise or modifying the model. We explore two scenarios: (1) both training and test data are encrypted, and (2) only training data is encrypted. Experimental results on a medical dataset demonstrate that the proposed framework maintains high prediction accuracy while significantly mitigating the risk of information leakage. This approach offers a practical and effective solution for protecting GP models in privacy-sensitive applications.