Research seriesL3algorithms
A zero-knowledge proof lets one party convince another that a computation was performed correctly without revealing the inputs — which is exactly what verifiable AI inference needs. This series builds the idea from interactive proofs and the simulator argument through Fiat-Shamir, SNARKs, transparent STARKs, and closes with zkML: proving a model produced an output on a committed input. Grounded in Goldwasser-Micali-Rackoff, Groth16, and Ben-Sasson STARKs.
A zero-knowledge proof convinces a verifier that a statement is true while revealing nothing else. The definition rests on three properties and one beautiful idea: the simulator.
Live back-and-forth is inconvenient. The Fiat-Shamir transform turns an interactive proof into a single string anyone can check by letting a hash function play the verifier's role.
A SNARK proves a huge computation was done correctly with a proof so small and quick to check it seems impossible. The price is a compilation pipeline and, often, a trusted setup.
STARKs drop the trusted setup entirely and rest on hash functions alone — which makes them transparent and plausibly quantum-safe, at the cost of larger proofs.
zkML applies zero-knowledge proofs to machine learning: a prover convinces anyone that a model produced a specific output on a committed input, without revealing the input or the weights.