Calibrating Federated Models for Reliable Mobile Edge Deployment

Abstract

Despite the rapid progress of federated learning (FL) for mobile and edge intelligence, the reliability of on-device federated models remains largely underexplored. This gap becomes critical in distributed mobile systems where clients are resource- and bandwidth-constrained, data are highly non-IID, and training suffers from intermittent participation and stragglers. In such settings, we systematically revisit FL from a reliability standpoint and identify an underappreciated failure mode: federated models often become poorly calibrated and unreliable in uncertainty estimation, which directly compromises risk-sensitive edge decisions. Through theoretical analysis, we further reveal that this degradation mainly arises from the classifier head, where biased local priors distort the global prediction confidence. To address this issue, we first propose Assembled Projection Heads (APH), a plug-and-play and communication-free head-ensemble framework that improves calibration by introducing controllable diversity while incurring negligible training and inference overhead. Building upon this, we develop Low-Rank Adaptation (LoRA)-based APH (LoRAPH), which replaces full projection heads with low-rank counterparts, effectively reducing storage overhead for large-scale distributed clients while further enhancing uncertainty reliability. Extensive evaluations across multiple benchmarks and FL frameworks demonstrate that APH and LoRAPH consistently improve calibration and robustness under non-IID data and partial participation, with minimal computation and communication overhead, making them practical for reliability-critical federated learning in distributed mobile systems. Notably, LoRAPH delivers reliable uncertainty estimation with only 0.34% extra training cost, near-zero inference overhead after pre-aggregation, and 4.37% footprint increase, making it well-suited for bandwidth- and memory-constrained mobile deployments.