Poster #54 - Bin Li

FSLearning: A Collaborative Deep Learning Framework for Efficient and Private Disease Prediction

Bin Li, M.S., Department of Computer and Information Sciences, Temple University, Philadelphia, PA, USA Xiaoqian Jiang, Ph.D., D. Bradley McWilliams School of Biomedical Informatics, UTHealth Houston, Houston, TX, USA Yu-Chun Hsu, Ph.D., D. Bradley McWilliams School of Biomedical Informatics, UTHealth Houston, Houston, TX, USA Arif O. Harmanci, Ph.D., D. Bradley McWilliams School of Biomedical Informatics, UTHealth Houston, Houston, TX, USA Hongchang Gao, Ph.D., Department of Computer and Information Sciences, Temple University, Philadelphia, PA, USA Xinghua (Mindy) Shi, Ph.D., Department of Computer and Information Sciences, Temple University, Philadelphia, PA, USA

Collaborative deep learning frameworks have become indispensable in healthcare, where sensitive biomedical data are distributed across institutions. Traditional federated learning (FL) enables multi-institutional model training without sharing raw data, but it remains vulnerable to privacy leakage from gradients and incurs substantial communication overhead. On the other hand, split learning (SL) enhances privacy by partitioning model architectures between clients and servers, yet it introduces high training latency due to its relay-based process. These limitations call for a new paradigm that can reconcile efficiency, accuracy, and strong privacy guarantees. We propose FSLearning, a federated split learning framework that combines the strengths of both FL and SL. It introduces three innovations: Tensor Regression Layers (TRLs): TRLs compress client activations before transmission, cutting communication by several orders of magnitude while maintaining accuracy within a few percentage points of FL. For example, on a ResNet3D backbone, parameter exchange drops from tens of millions to nearly one thousand per round with only a 0.02 accuracy difference. Flexible Privacy-preserving Mechanisms: FSLearning integrates either Homomorphic Encryption (HE) or Differential Privacy (DP), depending on application needs. TRLs enable efficient HE by reducing activation dimensionality, while DP achieves the lowest total variation distance (TVD ≈ 0.01) in membership inference evaluations, offering stronger resistance to privacy attacks compared to vanilla FL. Practical Efficiency and Scalability: Beyond accuracy and security, FSLearning significantly reduces computation time for multi-client training compared to FL. Its balanced design makes the framework especially well-suited for real-world biomedical imaging tasks, where both computational efficiency and privacy-preserving collaboration are essential. Through empirical evaluations, FSLearning demonstrates that it achieves comparable accuracy and communication efficiency to FL while improving scalability and privacy. Applied to disease prediction tasks, FSLearning emerges as a practical, secure, and efficient framework that addresses the core challenges of collaborative biomedical AI-balancing efficiency, adaptability, and privacy preservation.