In the case of AI fashions, one of many hardest inquiries to reply is deceptively easy: the place did this mannequin truly come from?
We addressed a part of this downside with Mannequin Provenance Equipment, an open-source instrument that fingerprints fashions on the weight stage (the parameters that defines what a mannequin is aware of and the way it behaves) to confirm their origins. However a fingerprinting instrument wants a transparent commonplace to measure in opposition to, that defines precisely what qualifies as a derivation relationship between two fashions. Right here, the business doesn’t but have a constant reply.
Definitions differ throughout licensors, requirements of our bodies, analysis teams, and AI labs. The identical pair of fashions will be labeled as “associated” by one reviewed and “unbiased” by one other, with each citing defensible reasoning. That inconsistency creates actual issues for licensing enforcement, vulnerability triage, and regulatory compliance.
We created the Mannequin Provenance Structure as an try to repair that. Comprised of a taxonomy, definition, and boundary specs, it is a normative reference, a structure, that specifies what a mannequin provenance relationship is and isn’t on the stage of weight derivation. This publish covers its construction, its reasoning, and the way it connects to the frameworks that governance packages already use. The Mannequin Provenance Structure builds on forthcoming work from Cisco AI Protection that describes the methodology in full, together with empirical proof for why such an method is important for each provenance and detection pipelines. You possibly can evaluate the Structure inside the docs folder of the Mannequin Provenance Equipment.
Why Defining Mannequin Provenance is Necessary
Basis fashions don’t arrive within the enterprise as remoted artifacts. They get fine-tuned, distilled, quantized, merged, and repackaged, and every step produces a brand new checkpoint whose relationship to its father or mother is poorly documented. When a safety group must know whether or not a deployed mannequin inherits a identified vulnerability, or when compliance wants to find out whether or not a third-party checkpoint triggers a licensing obligation, the query is at all times the identical: is that this mannequin a spinoff of that one?
And not using a shared, rigorous reply, group can face compounding dangers:
- Provide chain assaults are already exploiting this hole
- Regulatory necessities assume provenance readability that doesn’t but exist
- Incident response relies on traceable lineage
Provenance is About Mannequin Weights
The Mannequin Provenance Structure grounds provenance in a single idea: the verifiable derivation historical past of a mannequin’s educated weights. Two fashions share provenance if, and provided that, a causal chain of weight derivation connects them, whether or not instantly, not directly by distillation, or mechanically by a non-training transformation like quantization.
Shared structure, shared coaching knowledge, shared tokenizer, and shared benchmark efficiency don’t depend. The exclusion is deliberate. A broader definition that handled any architectural or behavioral similarity as derivation might make licensing enforcement apply to each mannequin in an structure household, would flag convergent designs as real vulnerability hyperlinks, and would flood governance audits with false positives. Weight-level causation produces labels which might be secure throughout reviewers, strong to metadata manipulation, and aligned with how derivation truly occurs in apply.
How Mannequin Provenance Structure is Structured
The structure solutions three questions: when are two fashions associated? How does that relationship happen? And what seems to be like a relationship, however isn’t? It organizes these solutions as express enumerations slightly than definitions-by-example, so each pair of fashions encountered in apply maps to a transparent class.
5 circumstances specify when a provenance hyperlink exists
- Direct descent: coaching initialized from a educated checkpoint
- Oblique descent: distillation from a trainer mannequin
- Mechanical transformation: quantization, pruning, merging, or format conversion
- Identification: byte-equivalent copy
- Transitivity: any composition of the above
A pair is provenance-linked if at the very least one situation holds.
9 mechanisms enumerate the concrete derivation pathways noticed in apply:
- Identification and reformatting
- Fantastic-tuning
- Continued pretraining
- Vocabulary-modified derivation
- Information distillation
- Structural modification with weight inheritance
- Quantization and compression
- Adapter-based derivation (LoRA, QLoRA, prefix tuning)
- Mannequin merging
Eight exclusions listed under are circumstances which will seem like provenance-linked, however are provenance-independent. Every exclusion is a sample of obvious similarity, however in the end one which carries no weight-derivation chain:
- Impartial copy (e.g., Llama-2 vs. Open LLaMA which share the identical structure and tokenizer, however are educated from scratch)
- Identical-family different-size (e.g., Llama-2-7B vs. Llama-2-13B).
- Identical-family different-corpus coaching (e.g., T5 vs. MT5, which share a reputation root, however have separate from-scratch coaching)
- Impartial runs below a shared seed (i.e., shared seed doesn’t represent shared weights)
- Architectural convergence (totally different groups independently arriving at comparable mannequin designs)
- Dimensional coincidence below totally different mechanisms (fashions that occur to share the identical dimension or form with out one being constructed from the opposite)
- Shared vocabulary with out weight switch (a tokenizer is a instrument, not a weight)
- Shared coaching goal (sharing an goal doesn’t hyperlink weights)
A rigorous provenance commonplace should title them explicitly, as a result of complicated any of them with real derivation corrupts downstream licensing choices, vulnerability assessments, and compliance determinations.
Establishing an Proof Customary
A taxonomy is barely as helpful because the proof commonplace connected to it. The Mannequin Provenance Structure accounts for 3 sources for establishing provenance (and however architectural similarity and naming conventions are explicitly inadequate):
- Official documentation: from the releasing group that explicitly names the father or mother mannequin and derivation technique
- Checkpoint verification: by hash matching, layer-by-layer comparability, or reproducible derivation scripts
- Authoritative third-party evaluation: that has been peer-reviewed or broadly cited
Beneath ambiguity, Mannequin Provenance Structure defaults to labeling a pair as provenance-independent. This conservatism is intentional. A false optimistic in provenance carries fast penalties: a licensing accusation, an IP declare, a supply-chain incident notification. A false adverse will get caught by defense-in-depth by guide evaluate, licensing audit, and forensic evaluation. Specificity wins when rigor is required.
Alignment with AI Risk Frameworks and Requirements
Mannequin provenance attestation will be thought of a provide chain management, and the Mannequin Provenance Structure serves as a definitional layer that makes mannequin dependency auditable. It specifies what it means for a deployed mannequin to inherit from an upstream supply, which is the precondition for any significant query about inherited vulnerabilities, license obligations, or unattributed redistribution.
“weak mannequin provenance” and noting that “no ensures on the origin of the mannequin.” The MITRE ATLAS framework paperwork provide chain compromise (AML.T0010) as a major preliminary-access method. The Cisco AI Safety and Security Framework classifies third-party mannequin elements below OB-009 Provide Chain Compromise, with direct applicability by AITech-9.3 (Dependency/Plugin Compromise). The Cisco AI Safety and Security Framework classifies third-party mannequin elements below OB-009 Provide Chain Compromise, with direct applicability by AITech-9.3 Dependency / Plugin Compromise: actors insert malicious code, backdoors, or vulnerabilities into third-party dependencies utilized by fashions, brokers, or AI purposes, creating supply-chain assaults that have an effect on all programs utilizing the compromised element. Basis-model checkpoints reused as initialization for downstream fashions are exactly such dependencies.
The structure additionally acknowledges the adversarial dimension by AITech-9.2 Detection Evasion: deliberate concealment of a derivation relationship — metadata rewriting, tokenizer substitution, chained modifications supposed to obscure the father or mother. The structure’s dedication to weight-level proof, slightly than metadata-level proof, is a direct response to this adversary mannequin.
Mannequin Provenance Structure attracts from current frameworks that AI provide chain packages already depend on. These frameworks determine necessities or issues that the structure helps fulfill. A proper provenance definition is a precondition for producing that documentation persistently throughout a corporation and throughout suppliers.
Desk 1. Frameworks, rules, and requirements that Mannequin Provenance Structure drew upon
A Dwelling Doc
New strategies of constructing fashions are rising quicker than any mounted taxonomy can accommodate. Mannequin merging, combining specialised educated fashions, has develop into a dominant method over the previous few years. Past merging, the ecosystem is seeing Combination-of-Consultants architectures with independently educated elements, federated coaching throughout organizations, and artificial knowledge pipelines that blur the road between data switch and unique coaching. The Mannequin Provenance Structure considers these open frontiers and commits to revision because the panorama evolves.
Get Began
The complete Mannequin Provenance Structure abstract is on the market alongside this publish: https://github.com/cisco-ai-defense/model-provenance-kit/tree/predominant/docs/structure
For groups able to put these definitions into apply, Mannequin Provenance Equipment offers the tooling. The whole pipeline runs on CPU, architectural matches resolve in milliseconds, and extracted options are cached for reuse. Try Mannequin Provenance Equipment Github: https://github.com/cisco-ai-defense/model-provenance-kit
Entry a starter set of base mannequin fingerprints on Hugging Face: https://huggingface.co/datasets/cisco-ai/model-provenance-kit
