The field of federated learning and reinforcement learning is moving towards addressing key challenges in data privacy and robust optimization. Researchers are developing novel methods to prevent data privacy leaks and improve the performance of federated reinforcement learning in the presence of mixed-quality data. Notably, new approaches are being proposed to adaptively adjust learning rates and identify high-return actions, leading to significant performance gains. Noteworthy papers include:

• A novel attack method that enforces prior-knowledge-based regularization to ensure reconstructed data remains close to the true transition distribution.
• A vote-based offline federated reinforcement learning framework that exploits a vote mechanism to identify high-return actions and alleviates the negative effect of low-quality behaviors.
• An adaptive decentralized federated learning approach that adaptively adjusts learning rates to mitigate the negative impact of abnormal clients on the global model.