Succinct zero-knowledge arguments (zk-SNARKs) allow a prover to convince a verifier of the truth of a statement via a succinct, efficiently verifiable proof without revealing any additional information about the witness. Despite their powerful capabilities and broad impact, practical deployments of zk-SNARKs remain limited to relatively small instances due to high proving/verification costs. With this motivation, we investigate methodologies that improve the concrete efficiency of zk-SNARKs and prepare them for large-scale, real-world deployment.

The first part of this talk focuses on applying zk-SNARKs to proofs of training (zk-PoTs). A zk-PoT enables an entity to prove that a committed model is faithfully trained on a committed dataset while the model and the dataset remain hidden from the verifier. We show how to realize efficient zk-PoTs for neural networks with optimal prover and verifier overheads.

In the second part, we study server-aided zk-SNARKs, which enable a prover to outsource most of its proving work to an untrusted server while the witness remains hidden from the server. We show how to achieve efficient server-aided proving for several widely deployed zk-SNARKs.