Verifiable, Private, Decentralized Computing in the Age of AI

Undergraduate

You type your question into ChatGPT and get back a response – how do you trust its accuracy? Perhaps you have reviewed the latest published benchmark results for GPT-4, or trust that others have. But how do you know the response you are getting from OpenAI’s servers is the true output of GPT-4? Perhaps due to a bug or system overload your question was handled by a weaker AI model. Or worse, perhaps the servers were hacked by someone maliciously giving you incorrect results. And, how do you trust that the sensitive questions you are sending to ChatGPT will not be leaked or used against you? This is a course in cryptographic proof systems. In the digital world today, we trust services to perform many kinds of computations on our data, from managing financial ledgers and databases to complex analytics, such as large-language model (LLM) inference. We trust these services not only to operate correctly, but also to keep our information private. Cryptographic systems allow us to remove this trust. A “succinct” cryptographic proof enables a service to attach a small certificate on the correctness of its computation which can be verified easily on a low-power device, even if the original computation was extremely complex. The proof of some ML computation that was run for hours on a GPU farm can fit in an email and take just milliseconds to verify on a mobile device! Beyond correctness, a “zero-knowledge” proof system enables us to prove knowledge of secret information, including hidden inputs to a computation that achieves a certain output. For instance, OpenAI could prove that the response is the true output of GPT-4 (a proprietary model) without revealing sensitive details about the model itself. In industry, the market for cryptographic proofs is currently around $75 million and projected to reach $10 billion by 2030, according to some estimates. Cryptographic proofs have become the leading technology for scaling blockchains and achieving privacy in cryptocurrencies. Verifiable and zero-knowledge computing also create an important foundation for decentralizing AI services. Training and serving large models currently require vast resources, leading to centralization. Decentralized ML networks offer a compelling alternative—letting many independent operators contribute incremental work to the overall task using their own resources big or small, from a single GPU to a server cluster, and earn a share of the payments clients make to use the service. In such a setting it is critical to verify that operators contribute incremental work correctly. Or, they may use private data to jointly train an ML model. Succinct zero-knowledge proofs would enable these operators to prove correctness of their work without revealing sensitive data. We will cover some of the other challenges and directions in decentralized model training and inference, such as reducing the amount of data that needs to be communicated between physically distributed islands of hardware and the potential role of reinforcement learning. Prerequisites: CPSC 201 and 202 (or equivalent, e.g. MATH 244). Recommended: CPSC 467 (Cryptography), MATH 225 (Linear Algebra). MW 1:05pm-2:20pm