Backdoor-Based Watermarking in Multi-Client Split Learning

Summary

This paper presents the first study of watermarking deep neural networks in multi-client split learning, where the model is partitioned between clients and a server. It proposes MarkSplit for benign settings and MarkSplit+ for settings with malicious clients, both using a watermark sample generation technique called Color-Shape-ID. Experiments report about 98% watermark detection accuracy with an average main-task accuracy drop of about 3%.