Federated Learning in Low-Resource Environments: Optimizing Local Models for Privacy-Preserving IoT Deployments

This paper presents a structured narrative review of optimization techniques for federated learning (FL) in low-resource Internet of Things (IoT) environments, with particular emphasis on privacy-preserving deployments. Although federated learning enables distributed model training without requiring raw data to leave individual devices, its practical implementation on resource-constrained IoT devices remains challenging due to limitations in computational capacity, memory, communication bandwidth, energy availability and heterogeneous data distributions. This review examines five interconnected dimensions of the problem: Computational and Memory Efficiency: Lightweight model architectures, weight quantization and TinyML-oriented training approaches. Communication Efficiency: Gradient sparsification, update compression, low-rank updates and communication-aware optimization. Statistical Heterogeneity: Approaches for addressing non-independent and identically distributed (non-IID) data across participating clients. Privacy Preservation: Differential privacy, secure aggregation and defenses against gradient inversion attacks. Energy Efficiency: Hardware-aware scheduling, adaptive federated learning rounds and energy-conscious client selection. The review synthesizes reported findings from existing research, highlighting the trade-offs between model accuracy, computational overhead, communication requirements, energy consumption and privacy guarantees. Particular attention is given to the limitations of existing evaluation practices, including the reliance on simulated environments, limited hardware-in-the-loop experimentation and the lack of integrated evaluations combining privacy, compression and statistical heterogeneity. The paper concludes by identifying research opportunities in joint optimization, hardware-grounded benchmarking, explicit threat modelling, resource-aware heterogeneity management and reproducible evaluation frameworks. Keywords: Federated Learning; Internet of Things (IoT); TinyML; Differential Privacy; Edge Computing; Model Optimization; Communication Efficiency; Resource-Constrained Devices; Privacy-Preserving Machine Learning.