Adaptive Multimodal Federated Optimization with Privacy and Blockchain Trust

Abstract

This article presents the Adaptive Multi-Modal Federated Optimization (AMMFO) framework which was developed to tackle important challenges in data privacy, fairness, and accountability in learning in the context of federated learning. The AMMFO framework addresses the need to train models with data from multiple modalities available at the edge in a secure and efficient manner while building trust and respecting privacy in a decentralized system. AMMFO applies differential privacy to protect sensitive model updates from adversarial inference and it leverages a blockchain-based trust mechanism to allow for transparency, immutability and decentralized accountability. The framework optimally adapts learning across multiple data modalities enhancing communication efficiency and stability of model convergence. Results from experiments indicate that AMMFO achieves 8-10% higher accuracy compared to FedAvg and FedProx, and 7-12% greater privacy resistance compared to FedDP while exploring different privacy budgets. Additionally, AMMFO improves convergence time by 15–20 percent and achieves less than 8% blockchain overhead. Overall, these results demonstrate the AMMFO framework's balance of performance, privacy, and scalability which enables next generation AI systems that are situated within privacy and trust-worthy frameworks in domains such as healthcare, finance, autonomous systems, and smart cities.