A Dynamic Federated Learning Framework with ...

A Dynamic Federated Learning Framework with Adaptive Differential Privacy for Cross-Institutional Artificial Intelligence Models

Author Name : Mohamed Ahmed Elmubarak Mohamed, Abulfathi Ibrahim Saleh Al-Hussaini, Hamid Basheer Almansor Adam, Rania Elsheikh Hamid Ibrahim

ABSTRACT Federated learning (FL) enables institutions to train a shared model without transferring raw records, but non independent and identically distributed data, gradient leakage, and network cost can reduce its practical value. This study specifies and evaluates an adaptive differential privacy federated learning framework (ADP-FL) that combines per-client gradient clipping, adaptive Gaussian noise, alignment-aware aggregation, and stochastic update quantization. The revised manuscript makes the evaluation protocol explicit: CIFAR-10 is partitioned among ten cross-silo clients with Dirichlet label skew, a ResNet-18 model is trained for 150 communication rounds, and each configuration is repeated over five seeds. Privacy is reported using Rényi differential privacy (RDP) accounting with sampling rate, clipping norm, noise multiplier, number of rounds, and δ stated explicitly. The supplied benchmark values show 96.45% accuracy, 95.80% precision, 94.60% recall, and an F1-score recalculated as 95.20%; round latency is reported as 1.12 s rather than sub-second. The manuscript also defines the Grad-CAM, quantization, and payload measurements required to support these claims. Because the original file did not contain raw logs or code, the numerical results should be rerun and independently verified before publication.