Federated Learning (FL) has presented one of the radical paradigms that permit collaborative models to be trained without centralized sensitive data, thereby supporting both privacy and compliance-related concerns. Nevertheless, issues involving communication bottlenecks, non-iid data distribution, and privacy leak channels continue to impede its general use. This research project puts forward an integrated approach integrating FL with Agentic Intelligence (AIgI), forming a decentralized, adapting framework that integrates self-governing decision-making capacity, dynamic coordination and enhanced optimized resources. The framework utilizes sophisticated FL optimization techniques, such as FedYogi, to improve convergence rate and model precision and the Communication-Aware Federated Learning (CA-FL) technique to minimize the bandwidth usage. An assessment of privacy risks is performed with the help of the FedInverse tool, which shows vulnerabilities to model inversion attacks and supports the idea of multi-layered defense mechanisms. As experimental results confirm, the proposed framework achieves excellent accuracy compared to conventional FedAvg, both in terms of precision, recall, and Dice similarity coefficient and scales to thousands of clients thanks to FEDn orchestration. The integration of agentic reasoning enables the system to be adaptive enough to cope with a heterogeneous environment, the lack of synchronous involvement of clients, and the trade-off between training during the optimization of network conditions. The presented work forms the basis of a new generation of privacy-preserving, decentralized AI architecture that is resilient, scalable, and apt to be used in sensitive applications, like healthcare, finance, and critical infrastructure.
Behera et al. (Fri,) studied this question.