Adaptive Privacy Budget Allocation for Differential Privacy Optimization in Fleet Federated Learning: Algorithm Enhancement and Performance Evaluation

Abstract

Federated learning in fleets faces critical challenges in balancing privacy protection with model performance when processing sensitive vehicular data. Traditional fixed privacy budget allocation strategies fail to account for the dynamic nature of distributed training across heterogeneous vehicle nodes. This research proposes an adaptive privacy budget allocation mechanism that dynamically adjusts differential privacy parameters based on training progression and parameter importance. The methodology integrates Fisher Information Matrix evaluation for layer-wise budget distribution and implements a round-based allocation strategy that concentrates privacy resources during critical learning phases. Experimental validation on the nuScenes and FEMNIST datasets demonstrates that the adaptive approach achieves 8.7% higher model accuracy than uniform budget allocation while maintaining equivalent privacy guarantees at ε=3.5. Communication efficiency is improved by 23.4% by reducing the number of convergence rounds. The framework provides fleet operators with practical guidance for implementing privacy-preserving collaborative learning systems that meet regulatory requirements while optimizing operational performance metrics.