A Privacy-Preserving Gaussian Process Regression Framework Against Membership Inference Attacks Using Random Unitary Transformation

As artificial intelligence (AI) systems become increasingly embedded in sensitive domains such as healthcare and finance, they face heightened vulnerabilities to privacy threats. A prominent type of attack against AI is the membership inference attack (MIA), which aims to determine whether specific data instances were used in a model’s training set, thereby posing a serious risk of sensitive information disclosure. This study focuses on Gaussian Process (GP) models, which are widely adopted for their probabilistic interpretability and ability to quantify predictive uncertainty, and examines their susceptibility to MIAs. To mitigate this threat, a novel defense mechanism based on Random Unitary Transformation (RUT) is introduced, which encrypts training and testing inputs using orthonormal matrices. Unlike Differential Privacy-based Gaussian Processes (DP-GPR), which rely on noise injection and often degrade model performance, the proposed method preserves both the structural integrity and predictive fidelity of the GP model without injecting noise into the learning process. Two configurations are evaluated: i) encryption applied to both training and test data, and ii) encryption applied only to training data. Experimental results on a medical dataset demonstrate that the framework significantly reduces the effectiveness of MIAs while maintaining high predictive accuracy. Comparative analysis with DP-GPR models further confirms that the proposed method achieves competitive or stronger privacy protection with less impact on model utility. These findings underscore the potential of structure-preserving transformations as a practical and effective alternative to noise-based privacy mechanisms in GP models, particularly in privacy-critical machine learning applications.

@article{Islam2025,
  title     = {A Privacy-Preserving Gaussian Process Regression Framework Against Membership Inference Attacks Using Random Unitary Transformation},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {16},
  number    = {8},
  year      = {2025},
  publisher = {The Science and Information Organization},
  author    = {Md. Rashedul Islam and Jannatul Ferdous Akhi and Takayuki Nakachi},
  doi       = {10.14569/IJACSA.2025.0160807},
  url       = {https://doi.org/10.14569/IJACSA.2025.0160807}
}