Weight diff
A comparison of a model checkpoint against its declared base model, weight by weight, to find changes that legitimate training would not produce.
Honest fine-tuning and quantisation change weights broadly and by small amounts; an implanted backdoor concentrates a large change in a handful of weight vectors. A weight diff is the one detection method Microsoft's Lazarus paper found reliable, but it requires access to the original weights, which is why provenance and private registries come first.
Related terms
DistillationFine-tuningModel provenancePost-trainingSleeper agent (backdoored model)
Get the brief in your inbox
One email every morning: the stories that mattered in AI, each with a one-line "why it matters" for your role. Free, no ads, unsubscribe any time.
Published daily by AI Briefing. We never share your address.