FedVCK: Non-IID Robust and Communication-Efficient Federated Learning via Valuable Condensed Knowledge for Medical Image Analysis

TL;DR

FedVCK tackles non-IID data in federated medical image analysis by condensing high-quality, necessary knowledge from each client into a small dataset using latent distribution constraints, and by updating the global model through relational prototype-wise contrastive learning at the server. The method combines distribution-matching condensation, latent-statistics fixing, and model-guided, importance-weighted data selection to minimize repeated knowledge and communication, while preserving privacy via random initialization of the condensed data. Empirical results across medical and natural datasets show FedVCK achieving state-of-the-art performance under non-IID and limited-communication regimes, with robust privacy properties. The approach offers a practical, scalable path for cross-institutional medical image analysis with strong non-IID robustness and communication efficiency.

Abstract

Federated learning has become a promising solution for collaboration among medical institutions. However, data owned by each institution would be highly heterogeneous and the distribution is always non-independent and identical distribution (non-IID), resulting in client drift and unsatisfactory performance. Despite existing federated learning methods attempting to solve the non-IID problems, they still show marginal advantages but rely on frequent communication which would incur high costs and privacy concerns. In this paper, we propose a novel federated learning method: \textbf{Fed}erated learning via \textbf{V}aluable \textbf{C}ondensed \textbf{K}nowledge (FedVCK). We enhance the quality of condensed knowledge and select the most necessary knowledge guided by models, to tackle the non-IID problem within limited communication budgets effectively. Specifically, on the client side, we condense the knowledge of each client into a small dataset and further enhance the condensation procedure with latent distribution constraints, facilitating the effective capture of high-quality knowledge. During each round, we specifically target and condense knowledge that has not been assimilated by the current model, thereby preventing unnecessary repetition of homogeneous knowledge and minimizing the frequency of communications required. On the server side, we propose relational supervised contrastive learning to provide more supervision signals to aid the global model updating. Comprehensive experiments across various medical tasks show that FedVCK can outperform state-of-the-art methods, demonstrating that it's non-IID robust and communication-efficient.

Figures (5)

• Figure 1: Illustration of the low synthesized data quality problem in Figure (a) and repeated knowledge problem in Figure (b).
• Figure 2: Overview of FedVCK. On the client side, we sample local data by importance sampling guided by the current model and then impose latent distribution constraints in optimization. We upload the condensed knowledge dataset and logit prototypes to the server. On the server side, we use cross entropy loss and relational contrastive loss to update the global model.
• Figure 3: We measure the average MMD of condensed knowledge class-wisely between adjacent rounds on the OrganC dataset. Greater MMD indicates a larger difference in distribution. The vanilla selection would cause more knowledge repetition between rounds. With our model-guided selection ($P_w$) in each round, the condensed knowledge between adjacent rounds is more different.
• Figure 4: Predictive performance under various $p$ values on medical datasets. 'Limited' indicates the performance under limited 10 communication rounds. 'Overall' indicates the overall performance with adequate communication rounds.
• Figure 5: Predictive performance under various $p$ values on natural datasets. 'Limited' indicates the performance under limited communication rounds (10). 'Overall' indicates the overall performance with adequate communication rounds.