Adaptive Client Selection in Tangle-Enabled Federated Learning for Efficient and Traceable Training

This paper proposes a robust and scalable Federated Learning (FL) framework integrated with the IOTA Tangle to address challenges of resource heterogeneity and participation instability in distributed environments. Traditional FL approaches often suffer from "stragglers" and lack a decentralized audit trail, leading to training inefficiencies. To mitigate these issues, we introduce an adaptive client selection mechanism driven by Reinforcement Learning (RL). Our approach employs a multifaceted Quality of Learning (QoL) metric that quantifies client contributions by evaluating update magnitudes, accuracy improvements, and hardware resource utilization. The RL agent integrates these QoL metrics into the reward function to dynamically optimize client subsets, ensuring faster con-vergence and efficient resource allocation. Furthermore, critical training metadata is encapsulated into IOTA Tagged Data blocks and anchored onto the Tangle, providing a tamper-resistant, traceable ledger. Experimental evaluations on the Flower platform demonstrate that our framework significantly improves training efficiency and provides a verifiable execution environment for edge intelligence.