Sept. 20, 2023, 1:10 a.m. | Lewis Birch, William Hackett, Stefan Trawicki, Neeraj Suri, Peter Garraghan

cs.CR updates on

Model Leeching is a novel extraction attack targeting Large Language Models
(LLMs), capable of distilling task-specific knowledge from a target LLM into a
reduced parameter model. We demonstrate the effectiveness of our attack by
extracting task capability from ChatGPT-3.5-Turbo, achieving 73% Exact Match
(EM) similarity, and SQuAD EM and F1 accuracy scores of 75% and 87%,
respectively for only $50 in API cost. We further demonstrate the feasibility
of adversarial attack transferability from an extracted model extracted via
Model Leeching …

accuracy attack chatgpt knowledge language language models large llm llms novel parameter similarity squad target targeting task

More from / cs.CR updates on

Business Information Security Officer

@ Metrolink | Los Angeles, CA

Senior Security Engineer

@ Freedom of the Press Foundation | Remote, 4 hour time zone overlap with New York City

Security Engineer

@ ChartMogul | Remote, EU

Malware Reverse Engineer

@ Two Six Technologies | Fort Meade, Maryland

SOC Analyst Level 3

@ OpenBet | Bengaluru, India

Course Developer, Network Security

@ Palo Alto Networks | Plano, TX, United States