Sampling-based, but on WEIGHTS, not iterations. CORRECTED 2026-07-13: this note used to say "proves a random subset of epochs/iterations rather than all of them" -- that is VeriML, not zkMLaaS. Per the ZKP-VML survey (our ONLY source; we hold no zkMLaaS PDF), zkMLaaS runs a two-round challenge-response in which the provider "submits commitments for all intermediate weights updated during each training iteration and data sampling epoch", after which "the ML client randomly selects a subset of these intermediate weights, and the service provider must generate corresponding proofs." The survey credits it with "approximately 273x reduction in proof-generation overhead" via random sampling + im2col + Freivalds-based matrix verification.