Secure Dynamic Spectrum Access in Internet-of-Things Based on Machine Unlearning

Dynamic spectrum access (DSA) has become a key technology to improve spectrum efficiency in Internet-of-Things (IoT) networks. However, traditional DSA methods require collection of user behaviour data and location information, hence posing concerns for leakage of privacy data from the AI models. In scenarios where strict privacy regulations require the deletion of privacy-sensitive user data, existing DSA mechanisms only have the AI models and without access to the training processes or training data. This paper proposes a DSA method for IoT by integrating Deep Q-Network (DQN) reinforcement learning with zero-shot machine unlearning. This novel approach addresses the dual challenges of efficient spectrum utilization and user privacy protection in IoT environments. By leveraging on knowledge transfer techniques in machine learning, our solution can effectively protect user privacy without directly accessing the original training data while maintaining the model's effectiveness on retained data. Furthermore, this method optimizes the learning process, ensuring that learning efficiency and model convergence speed are comparable to, and in some cases superior to, traditional reinforcement learning methods. Through a series of experimental validations, this paper demonstrates the effectiveness and efficiency of the proposal in IoT DSA scenarios. This research provides a new solution for IoT spectrum management and opens up new opportunities for privacy-enhanced machine learning applications.