Stealing Part of a Production Language Model | AI Paper Explained
Many of the top LLMs today are closed source. What if we could discover their internal weights?
In this video we dive into a recent research paper from Google DeepMind which presents an attack on large language models. The attack targets transformer-based LLMs, that expose log probabilities as part of their API, which includes GPT-4 and PaLM-2. The researchers successfully used the attack to discover internal data about OpenAI models. Part of the extracted data includes the hidden dimension size of gpt-3.5-turbo, and the researchers estimate it would take less than 2,000$ to extract the weights of the embedding projection layer of that model.
Stealing Part of a Production Language Model paper page - https://arxiv.org/abs/2403.06634
Blog post - https://aipapersacademy.com/stealing-...
-----------------------------------------------------------------------------------------------
✉️ Join the newsletter - https://aipapersacademy.com/newsletter/
👍 Please like & subscribe if you enjoy this content
We use VideoScribe to edit our videos - https://tidd.ly/44TZEiX (affiliate)
-----------------------------------------------------------------------------------------------
Chapters:
0:00 Introduction
1:13 Attack Targets
2:36 Hidden Dimension Extraction
5:29 Weights Extraction
6:18 Recover Logits From Log Probabilities
8:10 Results