Privacy-Preserving Agent-to-Agent Framework for Decentralized Multi-Modal Knowledge Sharing in Multimedia Learning Systems

This study proposes and empirically validates an Agent-to-Agent (A2A) decentralized and privacy-preserving learning framework to address the limitations of traditional Federated Learning (FL)—in-cluding data heterogeneity, high communication costs, server dependency, privacy leakage, and malicious participants. Unlike FL, A2A eliminates the need for a central server by enabling peer-to-peer exchange of distilled information (e.g., soft labels and trust scores) instead of model parameters, thereby mitigating data reconstruction risks and resolving the single point of failure (SPOF) issue. Experiments span five scenarios: non-IID data environments, communication efficiency, resistance to privacy attacks, robustness against adversarial agents, and multimodal learning scalability. In non-IID settings, A2A achieved a +1.3 percentage point accuracy gain; for communication, it showed reduced transmitted bytes per round; under gradient inversion attacks, the Attack Success Rate decreased from 0.579 (FedAvg) to 0.394 (−0.185 absolute). With 20% malicious participants, A2A reached 21.31% accuracy. In multi-modal experiments, A2A improved accuracy by 8% and cross-modal transfer by 2% over FL, albeit with a slightly lower F1-score. These results suggest A2A offers meaningful privacy protection and training stability in distributed environments and holds potential as a next-generation learning paradigm.