The Art of Mixology: Mixup-based Obfuscation for Privacy-Preserving Split Learning in Large Language Models
Researchers have introduced MIXGUARD, a new framework designed to enhance privacy in split learning for large language models (LLMs). This method employs token-level and representation-level obfuscation, along with adaptive gradient perturbation, to maintain learning signals while preventing data leakage to servers. Experiments demonstrate that MIXGUARD achieves utility comparable to non-split training and offers superior privacy protection against reconstruction attacks. AI
IMPACT Enhances privacy for LLM training by enabling split learning without significant utility degradation.