A Theoretical Comparative Study of Privacy Mechanisms and Optimization Techniques in Federated Learning for Edge AI

Federated Learning (FL) enables decentralized model training across edge devices while keeping raw data local, thereby enhancing privacy and reducing communication costs. However, theoretical challenges remain: reconciling model utility with formal privacy guarantees, mitigating communication and computation overheads, and ensuring robustness against adversarial participants. This paper presents a conceptual comparative study of prominent privacy mechanisms (differential privacy, secure aggregation, homomorphic encryption, and trusted execution environments) and optimization techniques (client selection, quantization and compression, adaptive learning, and personalization strategies) used in federated settings. Focusing exclusively on theoretical properties, we analyze the assumptions, security models, computational complexity, communication trade-offs, and utility-privacy relationships for each method. The goal is to provide a principled, design-oriented framework guiding researchers and practitioners for deploying FL in resource-constrained and privacy-sensitive environments.