Federated learning is a privacy-preserving machine learning framework that facilitates model training directly on decentralized data that, due to privacy concerns or transmission costs, cannot be centralized on a server for traditional model training. To prevent adversaries from reconstructing the original data via parameters transmitted during the process, homomorphic encryption is a commonly adopted method. However, it introduces significant communication and computation costs and risks total security failure if any secret key is compromised. This paper proposes a privacy-preserving aggregation mechanism that enables each client to independently generate partial keys for encryption while allowing decryption after homomorphic operations using an aggregated key. Key aggregation for the proposed algorithm is realized through secret sharing. Incorporating these components into a standard federated learning framework yields a novel method that enhances communication efficiency and offers robustness against privacy breaches from internal collusion. The algorithm’s resistance to linear and differential attacks is formally demonstrated by algebraically modeling the encryption procedure. Based on this analysis, the overall security of the method is likewise established. Experiments on the privacy-preserving aggregation mechanism demonstrate that the generated ciphertext exhibits favorable statistical properties and sensitivity. Simulation results of the federated learning method further indicate that, compared to existing encryption schemes, our proposed encryption method reduces communication cost by $77\mathrm{\%}\sim 93\mathrm{\%}$ with acceptable computational cost, thereby enabling lightweight encryption in federated learning.

This paper addresses the challenging problem of multi-agent dynamic target pursuit under stringent communication constraints (including delays and range limits), where the agile targets are non-cooperative and free from such limitations. To tackle this, we propose CRS-DQN, a novel Deep Q-Network algorithm designed for this scenario. CRS-DQN enables agents to learn effective pursuit strategies through deep reinforcement learning despite partial observability and constrained information sharing. Simulation experiments systematically evaluate the impact of key parameters. The results show that pursuit performance degrades monotonically with increased communication delay. In contrast, the communication radius exhibits a non-linear effect: performance peaks when the radius is within a specific range (approximately 1/10 to 1/5 of the environment size) and declines if the radius is too small or too large. Furthermore, an optimal balance exists between the communication radius and the delay threshold. This work demonstrates the feasibility of learning-based pursuit under strict communication constraints and provides insights into parameter tuning for robust multi-agent systems in adversarial, communication-degraded environments.