Hybrid Privacy Protection of IoT Using Reinforcement Learning

The smart mobile device, as an indispensable component of IoT, has been growing in volume and diversity in recent years. Along with this, cyber-physical social network (CPSN) experiences fast booming, in which users publish their posts or data for sharing. However, since the published data is usually public to all, adversaries can crawl data or launch attacks without much efforts. Existing research usually considers a static adversary where the attack is launched once to steal a type of sensitive information like identity or location. This is not practical in real-world scenarios. To release this assumption, we develop a hybrid privacy-preserving model that protects identity and location privacy at the same time against a dynamic adversary who actively launches attacks. In the proposed model, the privacy protection problem is also considered as a trade-off optimization model that users target on maximizing data utility with high-level privacy protection but adversaries try to achieve a opposite target. To make this happen, we model a multi-stage game built upon Markov Decision Process (GMDP). The user and the adversaries are regarded as two players (also known as parties) in this dynamic zero-sum game. The output of this game should be the optimal actions of users that an adversary cannot breach more privacy no matter how the actions changed. An improved reinforcement learning algorithm based on state-action-reward-state-action (SARSA) is developed, which reduces the cardinality from n to 2. This can significantly improve the convergence efficiency. At last, we conduct experimental simulations on real-world datasets to testify the superior performance on efficiency and feasibility over existing research. This paper is mainly based on our research on privacy protection using reinforcement learning [1–5].