The proliferation of data-generating edge devices and increasingly stringent data protection regulations such as the GDPR and HIPAA have made centralized machine learning training pipelines both impractical and legally risky. Federated Learning (FL) has emerged as a compelling paradigm that enables collaborative model training across distributed clients without requiring raw data to leave its point of origin. Despite this inherent privacy advantage, standard FL protocols such as Federated Averaging (FedAvg) remain vulnerable to gradient inversion and membership inference attacks, and further suffer from performance degradation under statistically heterogeneous (non-IID) client data. This paper proposes Differential Privacy-Enhanced Adaptive Federated Aggregation (DP-AFA), a framework that combines a calibrated Gaussian noise mechanism with a dynamic, quality-aware client weighting scheme to jointly address the privacy-utility-heterogeneity trilemma in federated systems. Local updates are perturbed under a formal (ε, δ)-differential privacy guarantee prior to transmission, while the server aggregates updates using weights derived from each client's local validation performance rather than raw sample counts alone. Experiments conducted on MNIST and CIFAR-10 under non-IID partitioning, benchmarked against FedAvg, FedProx, SCAFFOLD, and DP-FedAvg, show that DP-AFA achieves comparable or superior accuracy to non-private baselines while offering a substantially tighter privacy budget than standard DP-FedAvg, converging in 97 communication rounds versus 142 for FedAvg. These results indicate that formal privacy guarantees and competitive model utility are not mutually exclusive in federated settings when aggregation is designed adaptively.

Keywords — Federated Learning, Privacy-Preserving AI, Differential Privacy, Secure Aggregation, Non-IID Data, Distributed Machine Learning