US patent US12801792
Model checksums via deterministic graph traversals
Abstract
A computer-implemented method can include: capturing parameter data descriptive of values of parameters of a machine-learned model, the parameter data comprising a plurality of parameter lists respectively associated with a plurality of nodes of the machine-learned model, wherein the parameter data is captured during a forward-pass operation of the machine-learned model; sorting the parameter data to generate a sorted set of parameter lists; generating a plurality of parameter checksums, each checksum corresponding to one of the parameter lists of the sorted set of parameter lists; generating a model checksum of the plurality of parameter checksums; and authenticating the machine-learned model based on the model checksum.