Deep dives · 2026-11-19 · Advanced · By bombong
Model fingerprinting: tracing an LLM's origin from its weights and outputs
Model fingerprinting is a set of techniques for determining where a model came from using its weights, internal computations and output patterns. Approaches split into planting a mark in advance and reading features the model already has. Recent key issues are resistance to attacks and false judgments that arise in real auditing environments.
Contents
- Model fingerprinting identifies the origin of the model itself, which is a different goal from text watermarking, which traces generated text.
- Active fingerprinting plants a mark through fine-tuning, while intrinsic fingerprinting reads structure the model already has, such as its weight spectrum or layer-by-layer computation trajectory.
- Withstanding ordinary modifications (fine-tuning, quantization, merging) is a different problem from withstanding an adversary deliberately trying to erase the fingerprint. All 10 schemes studied were broken by adaptive attacks.
- Recent research goes beyond "are they related?" to the direction of "which one is the parent?", and addresses real auditing situations such as API resellers secretly switching models partway through a conversation.
- In practice, do not rely on a single fingerprint: combine white-box and black-box evidence and check the false positive rate separately.
Why model fingerprinting became necessary
LLMs (large language models) are expensive to build, and once released or leaked, anyone can modify the weights and redistribute them. The TrFP study paper points out that weights obtained without authorization are easily modified, which makes verifying their origin difficult. In the same vein, another study paper notes that as open-weight checkpoints are released and reused more often, model fingerprinting for ownership protection and post-hoc auditing has become important.
A common point of confusion here is how this differs from text watermarking. A survey on LLM copyright protection paper notes that existing reviews focused mainly on "techniques for tracing text a model generates," that is, text watermarking, while techniques for protecting the model itself (model watermarking and model fingerprinting) had not been systematically organized. The survey unified the terminology by placing model watermarking within the broader fingerprinting framework. This article follows that terminology.
The questions also take different forms. "Was this suspect model derived from ours?" (derivation detection), "Was this student model built by distilling our API outputs?" (distillation detection) and "Is the model the API promised really the one answering?" (identity verification) are each different questions. The available information also differs. If you can inspect the weights directly, it is a white-box setting; if you can only see queries and responses, it is a black-box setting.
Note that the sources for this article include no research on using weights to determine "what hardware a model was trained on." So this article covers only methods for verifying origin and derivation relationships.
Core idea: planted marks and innate fingerprints
By analogy, there are two branches. One is like engraving a serial number on a product. Planting a mark before shipping makes it easy to find later, but the engraving itself alters the product slightly, and someone could grind it off. The other is like reading features a person already has, such as fingerprints or irises. Nothing needs to be done in advance, but the features must be distinctive enough that similar people are not confused.
The first is active fingerprinting. As the survey paper and P2F paper explain, existing active methods plant an identity signal in the model through fine-tuning. A typical example is backdoor-based fingerprinting, which trains the model to give a fixed answer to specific questions (triggers). Accuracy and robustness are high, but P2F paper points out that each identity requires its own costly training, which limits scalability.
The second is intrinsic fingerprinting. It extracts features from the weights or internal computations without modifying the model. TrFP paper explains that this approach can protect ownership without modifying the model or embedding ownership information. In exchange, two requirements conflict. Fingerprints that compare weights directly break down easily when the model is modified (the robustness problem), and fingerprints that compare internal representations may judge unrelated models to be similar (the uniqueness problem).
Ultimately, design is a choice about "what to change in advance" and "what you can access." Active methods pay a cost up front and gain black-box verification. Intrinsic methods require no up-front work but generally need access to weights or internal computations. Black-box statistical methods that look only at outputs sit between the two and accept weaker evidence.
Three ways to obtain a fingerprint
ActivePlant a marker
- Changes the model, e.g. through fine-tuning
- Can verify models behind an API through queries
- Training cost for each identity
- Risk of quality loss and removal attacks
IntrinsicRead the built-in structure
- Does not change the model
- Requires access to weights and internal computation
- Trade-off between robustness and uniqueness
Output statisticsJudges from responses only
- Loosest access requirements
- Used to detect distillation relationships
- Sensitive to conversation history and decoding settings
공통 All three determine the source at the end with similarity scores or statistical tests
How it works
Active fingerprinting works in three steps. First, a mark is planted in the original model (via fine-tuning, watermark distillation, etc.), the model is deployed, and then queries are sent to a suspect model to check whether the mark appears. Watermark distillation embeds a statistical watermark signal in the weights so that even a model behind a black-box API can be identified paper. P2F paper changed the planting step. A dedicated generator takes text describing an identity and produces a low-rank parameter increment in a single computation, so no retraining is needed for each identity.
Intrinsic fingerprinting is divided by what it reads. The weight spectrum fingerprint paper summarizes the distribution of singular values (values indicating how much a matrix stretches in each direction) in the GATE projection weights of gated feed-forward layers. Layer-wise features are grouped by depth range into a model-level fingerprint. Because layers are not matched one by one, models of different depths can be compared. It needs no input samples, generated outputs or forward-pass activations, and involves no training. Rather than comparing weights or representations directly, TrFP paper treats how representations change across layers as a trajectory and builds the fingerprint from summary values that do not change when neuron order is permuted.
Multimodal models are more complicated. Because many models share the same language backbone, existing origin-verification methods are easily confused. AttnPrint paper uses only the low-frequency components of the cross-attention distribution between image and text as the fingerprint. Where only outputs are available, DistillTrace from the same study applies hypothesis testing to model outputs to find evidence of a distillation relationship.
Whatever the method, the final step computes the similarity between two fingerprints and compares it with a threshold to judge whether they are "from the same lineage." For example, the spectrum fingerprint computes similarity with linear CKA (a metric of how similar two representations are) paper. How to control false positives in this judgment is a key issue discussed later.
How the field developed: builders and breakers
A 2025 survey paper organized scattered techniques. It laid out the concepts running from text watermarking to model watermarking and fingerprinting, classified and compared fingerprinting techniques, and was the first to organize fingerprint transfer and removal techniques. As evaluation criteria, it proposed effectiveness, harmlessness (whether model performance is preserved), robustness, stealthiness and reliability. Later studies have largely aimed to push one of these criteria further.
Around the same time, research on the breaking side raised the bar. "Are robust fingerprints adversarially robust?" paper pointed out that existing evaluations covered only ordinary modifications such as further fine-tuning, model merging and prompt changes. The study set up a threat model assuming a malicious model host and broke 10 recently proposed fingerprinting schemes with adaptive attacks that fully bypass verification while preserving performance for users. A suppression attack study paper targeted ensemble settings where multiple LLMs are used together. Using TFA, which picks the next token only from a shared token set at each decoding step, and SVA, which filters out fingerprint responses using perplexity and voting, it suppressed backdoor fingerprint responses while preserving ensemble performance.
In 2026, research on the building side addresses these weaknesses from different directions. P2F paper tackles the cost problem of active methods, and the study revisiting watermark teachers paper addresses the trade-off between quality and detectability. That study points out that existing protocols underestimated quality loss in open-ended generation, which made overly strong watermarks look favorable. Because fingerprint verification can aggregate signals across many queries, the signal can be inserted more sparsely. It therefore proposed a near-tie restriction that applies the watermark only to tokens whose logit gap from the base model's top-1 prediction is small. On the intrinsic side, TrFP paper reported resistance to fine-tuning, pruning, quantization, model merging, parameter permutation and adaptive attacks, with fingerprint storage under 0.3 MiB. AttnPrint paper was tested on 154 model instances across 19 multimodal architectures and held up under five kinds of downstream modifications.
Where the field is heading
First, the question is shifting from "are they related?" to "which one is the parent?" Witness Overlap paper points out that existing origin verification mostly relies on symmetric evidence such as similarity in representations, weights or behavior. Knowing that A and B are similar does not tell you which came first. The study uses a third checkpoint from the same lineage as a "witness" and judges from the surrounding geometry which of the two candidates looks like the parent that the other branched from. Across 176 checkpoints from 16 lineages, a single witness was enough to get 95.3% of parent-child judgments right. This direction could be used to map family trees in an ecosystem where open-weight models are re-released after fine-tuning, alignment and merging.
Second, research addresses pitfalls that arise in real auditing environments. The contagious fingerprint study paper analyzed a situation where an API reseller answers the early part of a conversation with the model the customer paid for and secretly hands the rest to a cheaper model. The successor model imitates the style of the earlier conversation and inherits the previous model's fingerprint. As a result, a text classifier's detection rate for the substitute model dropped from 0.97 to 0.72 when DeepSeek replaced Claude Sonnet, and from 0.87 to 0.54 when downgraded to Haiku from the same company. Conversely, honest models also drift from their own fingerprint in long conversations. An active probe calibrated on fresh conversations flagged honest models in long contexts at rates of up to 1.00. The study diagnoses both failures as coming from comparing "answers that depend on conversation history" against "history-free baselines." Its alternative, Lineup, samples baseline answers from the claimed model with the same conversation history and tests where the audited answer ranks among them. On two closed models, the false positive rate was 0.056.
Both studies send the same message: high accuracy under lab conditions does not guarantee discriminative power in real audits, and designing the comparison baseline to fit the auditing situation matters.
Notable numbers from recent research
Limitations and open issues: what to weigh in practice
The biggest issue is the threat model. Results showing resistance to fine-tuning or quantization concern cases of "unwitting modification." The conclusion of the adversarial robustness study paper is that a different standard is needed against an adversary who knows the fingerprinting scheme and tries to erase it, and the study recommends considering adversarial robustness from the design stage. Backdoor fingerprints can be suppressed in setups that filter responses, such as ensembles paper. When reading an evaluation report, first check what kind of attacker was assumed.
The second is trade-offs. Active fingerprinting modifies the model, so it carries a harmlessness cost, and stronger watermarks make detection easier but degrade text quality paper. Intrinsic fingerprinting does not modify the model, but often requires access to weights or internal computations. That makes it hard to apply directly to models hidden behind an API. Output-based methods have the loosest access requirements but are affected by the environment, such as conversation history and decoding settings paperpaper.
The third is false positives. If similarity thresholds are set empirically, honest models may come under suspicion when conditions change. Where outcomes affect people and organizations, such as ownership disputes or judgments of contract violations, report the false positive rate separately and do not draw conclusions from a single fingerprint result.
There are three common mistakes: using a text watermark detector as-is as evidence of model origin; comparing only the language part of multimodal models that share the same backbone and judging them to be "the same model" paper; and concluding the direction of derivation solely because similarity is high paper.
Checks before trusting fingerprint results
Threat model
- Were only ordinary modifications tested, or adaptive attacks as well?
- Were ensemble and response-filtering setups considered?
Access requirements
- Is weight access available, or only an API?
- If multimodal, could a shared backbone cause confusion?
Judgment
- Was the false positive rate checked separately?
- Were thresholds set for the audit conditions (conversation history, etc.)?
- Was the direction of derivation not concluded from a similarity result alone?
Further reading
If you are new to the topic, start with the survey paper to get the terminology and evaluation criteria down. Once the relationship between text watermarking, model watermarking and model fingerprinting is clear, you can see where later papers fit. Next, reading the adversarial robustness study paper shows how to look critically at claims of being "robust."
To go deeper by approach: for active methods, read P2F paper and the study revisiting watermark teachers paper to see the two axes of cost and quality. For intrinsic methods, read the weight spectrum fingerprint paper and TrFP paper together, and compare how the choice of comparing weights, representations or trajectories changes robustness and uniqueness. For multimodal models, see AttnPrint and DistillTrace paper.
If you are interested in real auditing scenarios, we recommend the contagious fingerprint study paper and Witness Overlap paper. The former covers the pitfalls of API identity verification, and the latter covers determining the direction of open-weight lineage.
Related terms
Open weights AI watermark Distillation Backdoor attack Model extraction AI supply chain security Model merging Checkpoint
References
Foundational papers
- Copyright Protection for Large Language Models: A Survey of Methods, Challenges, and Trends (2025)
- Prompt2Fingerprint: Plug-and-Play LLM Fingerprinting via Text-to-Weight Generation (2026)
- Language Model Fingerprinting Requires Rethinking Watermark Teachers (2026)
- Fingerprinting Multimodal Large Language Models (2026)
- From Representations to Computational Trajectories: Intrinsic Fingerprinting for Large Language Model Provenance Verification (2026)
- Model-intrinsic, Weight-level, and Training-free Spectral Fingerprinting for Large Language Models (2026)
- Are Robust LLM Fingerprints Adversarially Robust? (2025)
- Inhibitory Attacks on Backdoor-based Fingerprinting for Large Language Models (2026)
Recent research
Written by AI from the listed references and reviewed. Please check the original sources for details.