Edge AI and Federated Learning for Privacy-Preserving IoT Systems: System Architectures, Non-IID Optimization, and Security Bounds

Abstract—The proliferation of Internet of Things (IoT) devices has created an unprecedented surge in decentralized telemetry, sensor data, and localized computational workloads. Conventional cloud-centric machine learning architectures require centralized pooling of raw edge data, which induces prohibitive wide-area network (WAN) bandwidth consumption, unsustainable latency, and acute privacy vulnerabilities under international compliance mandates (e.g., GDPR, HIPAA). The convergence of Edge AI and Federated Learning (FL) establishes an alternative computational paradigm where edge nodes perform local model optimization on private data, communicating solely parameterized weight or gradient updates to an aggregation coordinator. Nevertheless, deploying robust FL across heterogeneous IoT ecosystems introduces profound systems challenges: non-independent and identically distributed (non-IID) data distributions, severe hardware resource disparities (memory, compute, power), communication bottlenecks over constrained wireless channels, and susceptibility to model inversion, gradient reconstruction, and Byzantine poisoning attacks. This paper delivers a comprehensive technical survey and systematic architectural formulation of privacy-preserving Edge-FL for IoT systems. We formalize the distributed optimization objective under statistical skew, analyze parameter-efficient edge acceleration techniques (including dynamic quantization and parameter-efficient fine-tuning), and evaluate cryptographic privacy mechanisms combining Differential Privacy (DP) and Secure Multi-Party Computation (SMPC). Furthermore, we conduct extensive empirical benchmarks across 100 heterogeneous edge nodes simulating industrial IoT, smart healthcare, and connected mobility workloads. Our results demonstrate that adaptive quantized federated aggregation achieves an 81.4% reduction in uplink communication bandwidth while preserving model accuracy within 1.6% of centralized benchmarks under severe statistical skew. Finally, we formulate open challenges in post-quantum cryptographic federated verification, asynchronous edge coordination, and green zero-carbon IoT optimization. Index Terms—Edge AI, Federated Learning, Internet of Things (IoT), Differential Privacy, Non-IID Optimization, Parameter-Efficient Fine-Tuning, Secure Multi-Party Computation, Edge Acceleration.