Stealing LLMs (MIT, Microsoft, Harvard) #ai

Reverse-Engineering LLMs through Conditional Queries and Barycentric Spanners. Excellent new AI research by MIT, regarding cybersecurity and data privacy. MIT published today a new method for efficiently learning and sampling from low-rank distributions over sequences, a problem central to understanding and replicating the behavior of complex language models (LLMs). Novel approach assumes access to an unknown distribution H over sequences of length T from an alphabet O, where H has a rank S due to its low-dimensional underlying structure. By leveraging conditional queries to H, the authors (MIT) construct a compact representation that captures the essential features of the distribution without exhaustively exploring the exponential sequence space. Key to the new mathematical method is the use of barycentric spanners and dimensionality reduction techniques. Identify a small set of representative histories H(t) whose conditional distributions can approximate all other histories through bounded linear combinations. By constructing succinct representative vectors and employing convex optimization to minimize the Kullback-Leibler divergence during the sampling process, this new method ensures that errors remain controlled and do not compound significantly. This allows to reconstruct the transition model implicitly and generate sequences that closely mimic those produced by H, effectively reverse-engineering the LLM's behavior while maintaining computational efficiency. And all of this without knowing the parameters and the training data of the unknown LLM. AI Mathematics is amazing. All rights w authors of arXiv pre-prints: ------------------------------------------------------------ Model Stealing for Any Low-Rank Language Model @mit https://arxiv.org/pdf/2411.07536 YOU HAVE TO read this excellent pre-print before: Learning Hidden Markov Models Using Conditional Samples @harvard @MicrosoftResearch @UCSanDiego https://arxiv.org/pdf/2302.14753v2 00:00 Model Stealing for ANY Low Rank Language Model 00:35 Learning Hidden Markov Models 02:00 Reverse-Engineer LLMs 03:28 Professor of Mathematics MIT 06:08 Hidden Markov Models explained 09:13 New method 11:37 Barycentric Spanner explained 14:37 Convex Optimization KL Divergence 19:12 Low Rank Distribution explained 20:28 MAIN Challenge 25:37 The MAIN Mathematical Theorem #education #science #mathematics