The convergence of artificial intelligence (AI) with decentralized computing has introduced a new generation of distributed intelligent systems operating without centralized authority. This paper presents an extended investigation of the architectural foundations, threat landscape, and mitigation strategies of decentralized AI systemsTo understand how federated learning, blockchain, and edge computing converge, this paper introduces a structured taxonomy based on four core dimensions: architecture, trust, security, and governance. Key attack vectors, including model poisoning, Sybil attacks, gradient inversion, and consensus manipulation, are analyzed alongside defenses such as robust aggregation, differential privacy, and lightweight cryptographic protocols. A comparative analysis of representative frameworks further highlights trade-offs between performance, scalability, and security. Literature-based observations show that hybrid blockchain-FL architectures can achieve accuracy above 90% while reducing communication overhead by up to 60%, though security guarantees remain scenario-dependent. To guide future work, the paper highlights critical open challenges in standardization, adaptive trust, and real-world deployment.