Backdoor-Based Watermarking in Multi-Client Split Learning
Summary
Split learning (SL, a technique where a deep neural network is divided between a client's local computer and a server to reduce computation on the client side) faces challenges in protecting intellectual property through watermarking (a hidden mark added to prove ownership) in multi-client settings, because the server can erase watermarks, later clients can overwrite earlier ones, and malicious clients can deliberately remove them. This paper proposes MarkSplit and MarkSplit+, two methods that embed watermarks more robustly by jointly training the main task with watermark samples in a three-tiered training structure, with MarkSplit+ using dynamic adjustment for adversarial environments with malicious participants.
Solution / Mitigation
The source proposes two explicit methods: (1) MarkSplit for benign environments, which jointly trains main-task and watermark samples within a three-tiered structure (mini-local, local, and global rounds); and (2) MarkSplit+ for adversarial settings, which enhances robustness by dynamically adjusting watermark sample counts per client based on watermark detection accuracy. Both use a watermark sample generation technique called Color-Shape-ID.
Classification
Related Issues
Original source: http://ieeexplore.ieee.org/document/11602638
First tracked: September 3, 2026 at 08:02 PM
Classified by LLM (prompt v3) · confidence: 85%