Federated Agentic Learning Algorithms for Privacy Preserving Collaborative Intelligence

Abstract

Data privacy concerns are often a constraint for collaborative intelligence in distributed systems, as is the lack of agent autonomy. In this paper, we present a Federated Agentic Learning (FAL) framework that combines the strengths of autonomous agent decision-making and federated learning to achieve collaborative intelligence while preserving privacy. Each agent uses policies based on reinforcement learning to maximize the local actions, engages in federated aggregation of model updates, and incorporates differential privacy and secure communication protocols. The framework was tested on three datasets, representing healthcare, manufacturing, and finance data, with non-IID, heterogeneous data distributions. Experiments show that FAL outperforms the standard federated learning model and the decentralized model without an agent in accuracy (92.3%, 94.0%, 91.8% on the three datasets) and F1 score (91.1%, 93.5%, 91.2% on the three datasets), while improving the communication efficiency and privacy protection (p < 0.05). The results show that agentic policies can improve the convergence and robustness of the models whilst preserving the confidentiality of sensitive data. The proposed framework is scalable and secure and can be used in various domains such as financial networks, industrial IoT, and healthcare networks. In the future, dynamic agent selection, adaptive privacy budget, and edge deployments for large-scale heterogeneous environments will be explored.