Optimizing Privacy and Utility in Statistical Analyses using Multi-Armed Bandits : AI enabled data privacy protection

Sammanfattning

In the digital era, the exponential increase in data generation has greatly amplified the importance of statistical analysis for informed decision-making across sectors like healthcare, economics, and technology. However, big data introduces significant privacy challenges, with sensitive information at risk of inference attacks, leading to potential privacy breaches and General Data Protection Regulation (GDPR) violations. This thesis explores the intersection of statistical data analysis and data privacy, focusing on methods to protect sensitive data from inference attacks while balancing privacy and utility. This balance is governed by hyperparameters, traditionally tuned manually, which is time-consuming and suboptimal. To address this problem, this thesis employs Reinforcement Learning (RL), specifically Multi Arm Bandits (MABs), to optimize hyperparameters for effective privacy-utility trade-offs. The thesis introduces ”PrivaGym”, a novel framework designed to identify optimal hyperparameters for data privacy protection methods, maximizing both privacy and utility. PrivaGym systematically explores hyperparameter configurations of privacy protection mechanisms, leveraging MAB to achieve the best possible balance between privacy and utility. Towards this goal, PrivaGym simulates inference attacks, applies privacy protection methods with various hyperparameters, and quantifies the privacy gain and retained utility. The MAB agent interacts with PrivaGym’s environment to ascertain the optimal hyperparameters that yield maximum privacy gain and utility. This study evaluates PrivaGym by employing defenses based on Differential Privacy (DP) for privacy protection against Membership Inference Attacks (MIAs) and Reconstruction attacks, and various utility metrics for statistics like sum, mean, and standard deviation, on synthetic and real-world datasets. Overall, PrivaGym is a robust framework that ensures data privacy while maintaining utility in various settings, facilitating informed decision-making and strategic planning in the data-driven world, and contributing significant insights to data privacy.