PP-CLS-FL: Privacy-Preserving CLS-Transfer Federated Learning on Heterogeneous Clients

Federated learning (FL) is a decentralized paradigm that enables collaborative model training across multiple institutions—such as hospitals—without sharing sensitive medical data (e.g., MRI or CT scans). The most common algorithm, Federated Averaging (FedAvg), aggregates locally trained model weights to form a global model. However, FedAvg struggles under statistical heterogeneity (non-independent and identically distributed data from diverse scanners, imaging protocols, and patient demographics) and system heterogeneity (varying client computation and memory capacities), often leading to unstable optimization and slow convergence.To address these challenges, we propose CLS-Transfer Federated Learning (CLS-TFL), a framework that improves robustness and efficiency in heterogeneous medical FL. Our approach leverages a pretrained Vision Transformer (ViT) while adapting only the [CLS] (classification) token head—and optionally the final transformer block—keeping the main encoder frozen. This design reduces trainable parameters and stabilizes non-IID training by focusing adaptation on the [CLS] representation that summarizes global image semantics.We further introduce a resource-aware orchestrator that estimates client training speeds during a warmup phase and dynamically allocates local epochs to satisfy a time budget while ensuring adequate sample coverage for rare data types. Experiments on a non-IID MRI classification dataset with client-specific perturbations (blur, noise, brightness, resolution) demonstrate that FedAvg achieves 73.56 % test accuracy, whereas CLS-TFL reaches 89.49 %. Incorporating resource-aware allocation further improves accuracy to 92.29 % while reducing average round time from 66.10 to 62.40 seconds.Unlike existing optimizer- or client-selection-based approaches, CLS-TFL unifies ViT [CLS]-based adaptation with time-budgeted orchestration, improving both accuracy and time-to-accuracy under real-world heterogeneity. The framework is modular, easily deployable, and compatible with personalization and fairness extensions in medical federated learning.