Introducing RMIA: A New Privacy Risk Analysis Method for Machine Learning
Privacy in machine learning models has become a concern due to Membership Inference Attacks (MIA). These attacks check if specific data points were part of a model’s training data. Understanding MIA is critical as it assesses potential exposure of information when models are trained on different datasets. MIA’s scope includes statistical models, federated learning, and privacy-preserving machine learning.
MIA methods have evolved from summary statistics to hypothesis testing strategies and approximations, especially in deep learning algorithms. Previous MIA approaches have faced challenges due to the computational demands of privacy audits.
A new paper proposes a novel approach to Membership Inference Attacks (MIA) called Relative Membership Inference Attack (RMIA). This method aims to discern whether a specific data point was used during training of a machine learning model.
RMIA introduces a refined likelihood ratio test that measures the distinguishability between data points. It leverages population data and reference models to enhance attack potency and robustness against adversary background knowledge variations. RMIA consistently outperformed other attacks across various scenarios and datasets.
RMIA emerges as a robust, high-power, cost-effective attack, outperforming prior state-of-the-art methods. It showcases superior performance with limited reference models and provides reliable and adaptable membership inference attacks for privacy risk analysis tasks in machine learning models.
To find out more about this research, click here. Follow us for the latest AI research news and cool AI projects. Check out our website for more information.