PrivaGym 2.0: Automating Privacy Defense Strategies with Reinforcement Learning.

This thesis addresses the problem of balancing privacy and usefulness in sharing data for machinelearning (ML) analytics, cybersecurity, and scientific research. Existing privacy-protectivemeasures do not offer sufficient protection without sacrificing data utility. This research solves thisissue by using reinforcement learning (RL) to develop an independent agent capable of learningdynamic privacy-preservation strategies in an adversarial environment. With its interactions withprivacy attackers and data evaluators, the RL agent learns to optimize data transformations tosecure sensitive information while leaving its analytical usefulness undeterred.Based on this idea, the thesis presents an RL-based modular framework that embeds multipleprivacy-preserving defense tools, such as pre-processing, noise injection, sampling, aggregation,transformation, and post-processing into an RL-based setting. These tools are parts thatcan be used to create strong privacy-preserving systems rather than stand-alone defenses. Theenhanced version of the framework further enhances the approach by taking advantage of statisticalsummaries for better state representation, broadening the space of actions to includevarious privacy-protective tools and algorithms, and allowing for context-sensitive and adaptableRL agent to make subtle decisions in various scenarios. Additionally, this framework emulatesinference attacks in real-world scenarios, such as Membership Inference, Reconstruction, andLinkage Attacks, and utilizes adversarial as well as utility feedback for learning. Multiple RL algorithms,Proximal Policy Optimization (PPO), Deep Q-Networks (DQN), Trust Region Policy Optimization(TRPO), and Advantage Actor-Critic (A2C), are tested for efficacy in learning optimalstrategies in various privacy-utility trade-offs. The work further examines performance across differentdatasets and configurations, providing insight into adaptively improving data privacy withoutlargely compromising analytical performance.By developing and validating this novel framework, this thesis intends to contribute to state-of-theartprivacy-preserving technologies and enable secure, context-adaptive data sharing in practicalscenarios wherein protecting privacy and efficiency in utility are equally essential.