Exploring the Role of Zero Knowledge Proofs in Large Language Models

The recent swift emergence of artificial intelligence (AI) and the pioneer introduction of Large Language Models (LLM) in particular, has brought change and disruption to most industries, allowing achievements in the areas of text generation, text summarization, and automated reasoning to be leap-frogged. Nevertheless, such models use huge data sets, most of which are proprietary in nature, and as such, casts serious concerns over privacy, security, as well as trust Since the weighting of the LLM is usually stored as intellectual property, it is problematic that their results should be directly verified. This transparency problem has brought up both legal and ethical concerns of the integrity and reliability of their outputs zkLLM is based on tlookup, a parallel lookup argument specifically designed to prove LLM computation without disclosing sensitive model information [11]. This invention brings in no new overhead but maintains computational efficiency. It is based on this that we suggest zkAttn, a zero-knowledge proof of the attention mechanism, which trades off on speed, memory footprint, and precision. With our full parallelization on CUDA, it is now possible to have zkLLM produce proofs of 13 billion parameters LLM in under 15 minutes, and generate efficient proofs, less than 200 kB in size, everything under the covers of secretly hold onto the model personal information. Whereas checking of LLM outputs can be considered a highly crucial operation, the larger issue is verifiable machine learning (VML) with particular focus on outsourced or federated learning where the calculations are done by third parties.