Reputation-Based Peer Selection in Decentralized Networks Using Distributed Federated Learning Models

Decentralized networks face significant challenges in peer selection due to the absence of centralized trust mechanisms, leading to vulnerabilities from malicious nodes and suboptimal resource utilization. Traditional federated learning approaches struggle with data heterogeneity and Byzantine attacks, compromising model integrity and convergence efficiency. This study proposes a novel reputationbased peer selection framework integrated with distributed federated learning models, employing dual-reputation computation schemes (debit-credit and credit-only) to evaluate peer contributions objectively. The system implements a layered architecture combining blockchain-based reputation storage with adaptive aggregation algorithms, enabling dynamic peer ranking and selective participation in model training rounds. Experimental validation demonstrates significant improvements in model accuracy (94.7% vs. 87.2% baseline), Byzantine fault tolerance (withstanding up to 35% malicious nodes), and communication efficiency (42% reduction in network overhead). The reputation system achieved 96.3% accuracy in malicious peer detection while maintaining 89.1% model convergence rate under heterogeneous data distributions. The proposed framework effectively addresses trust and security challenges in decentralized federated learning environments, providing robust peer selection mechanisms that enhance overall system performance and reliability while preserving data privacy and enabling scalable distributed machine learning applications.