mimoo/awesome-zk-ai

This repository is a curated knowledge base about making AI systems trustworthy, built and maintained by the security research group zkSecurity. Despite the zk in its name, zero-knowledge proofs are just one of several techniques it covers, not the whole focus. The main way to browse it is a generated website with a dedicated page for every approach and every paper, where every number shown is pulled directly from the underlying data along with where that number came from. The project organizes the field around two questions: which phase of a model's life you are protecting, either when it is running (inference) or when it is being trained, and which property you actually want, either proof that something was done correctly (verifiability) or protection of the data and model while computation happens (privacy). Each of those four combinations is covered by a different set of cryptographic and non cryptographic techniques, including zero-knowledge proofs, secure multi-party computation, fully homomorphic encryption, and trusted hardware, each with very different cost and trust tradeoffs. Verifiability proves correctness to a third party but is roughly a thousand times more expensive computationally, while privacy hides data from a counterparty at the cost of heavy communication or encrypted computation. Within the verifiability side specifically, the project further splits things by what exactly is being proven, from the cheapest claim that a specific output came from a specific input, up through proving a model's accuracy on a dataset, and the most expensive claim that a model was trained correctly on specific data. There is also a landscape page mapping out which companies and research groups are actually building in each of these areas, noting that genuinely focused zkML companies are a small group, while most other activity comes from broader cryptography, encryption, or decentralized AI projects adding related language to their existing work.