Performance and Information Leakage in Splitfed Learning and Multi-Head Split Learning in Healthcare Data and Beyond.

Area of Science:

• Healthcare Data Analytics
• Machine Learning
• Distributed Learning

Background:

• Machine learning (ML) offers powerful insights for healthcare decision-making.
• Stringent data privacy laws necessitate privacy-preserving ML techniques.
• Split-Fed Learning (SFL) enables collaborative ML model training while protecting data privacy.

Purpose of the Study:

• To investigate a modified SFL architecture, Multi-Head Split Learning (MHSL), that removes client-side model synchronization overheads.
• To evaluate the trade-offs between computational efficiency, model accuracy, and information leakage in MHSL compared to SFL.
• To analyze the impact of client-side model depth on performance in both SFL and MHSL.

Main Methods:

• Implemented and evaluated SFL and MHSL architectures.
• Utilized Resnet-18 and Conv1-D models on ECG and HAM-10000 datasets.
• Assessed model performance (accuracy) and information leakage using mutual information scores.
• Varied the depth of the client-side model portion to analyze its impact.

Main Results:

• SFL achieved higher accuracy than MHSL by 1.81% (ECG) and 2.36% (HAM-10000) at a cut-layer of 1.
• Increasing client-side model layers improved SFL performance but degraded MHSL performance.
• SFL exhibited greater information leakage than MHSL on both datasets.

Conclusions:

• MHSL offers reduced computational overhead but sacrifices accuracy and privacy compared to SFL.
• Client-side model architecture significantly influences the performance of both SFL and MHSL.
• Careful consideration of the trade-off between efficiency and accuracy/privacy is crucial when selecting between SFL and MHSL for healthcare applications.