Multi-Agent Energy Trading with Privacy Heterogeneity: A Denoise Dynamic Differential Privacy Multi-Agent Reinforcement Learning Method

With the rapid development of multi-agent collaborative decision-making in smart grids, electricity trading systems increasingly rely on large-scale user load data to optimize trading strategies. However, heterogeneous agents exhibit diverse privacy requirements, while existing unified privacy-preserving mechanisms struggle to balance personalized privacy demands with overall system efficiency. To address this challenge, we propose a dynamic differential privacy scheme tailored for multi-agent game environments. Specifically, we first design a personalized privacy risk assessment metric based on weighted mutual information to quantify the privacy leakage risk of diverse users. Then, we formulate the electricity trading problem as a multi-objective optimization that captures the trade-off between utility and privacy, and establish a Markov decision process model to maximize system economic efficiency under privacy-preserving constraints. Finally, we develop an adaptive noise-injection reinforcement learning framework integrated with a denoising network, and theoretically prove that the proposed noise mechanism satisfies ε-differential privacy. Simulation results demonstrate that, compared with benchmark methods using fixed-noise and personalized-noise mechanisms, our approach enhances privacy protection effectiveness by 79.11% and 77.76%, respectively. Moreover, the incorporation of the denoising network eliminates 50.97% of the cost increase induced by noise injection, thereby achieving a coordinated optimization of privacy preservation and operational efficiency.