New AugMP strategy targets federated fine-tuning of LLMs

Researchers have developed a new strategy called Augmented Model Manipulation (AugMP) to counter threats in federated fine-tuning (FFT) of large language models (LLMs). AugMP utilizes a graph representation learning framework to identify correlations in legitimate LLM updates, which then guides the creation of malicious updates. An iterative algorithm optimizes these adversarial updates to embed malicious objectives while maintaining a benign appearance, making them difficult to detect by standard defense methods. Experiments show AugMP can significantly reduce the accuracy of global LLMs by up to 26% and local agents by up to 22%. AI

IMPACT Introduces a novel attack vector against federated LLM training, highlighting the need for advanced defense mechanisms.

RANK_REASON Research paper detailing a new method for model manipulation in federated learning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• alphaXiv
• arXiv
• AugMP
• DagsHub
• federated learning
• Gotit.pub
• Hanlin Cai
• Hugging Face
• LLMs
• ScienceCast

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →