- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This paper proposes scalable semantic fingerprinting for black-box verification of LLM ownership. A fingerprint unit consists of a seed concept, a family of 20 paraphrased questions, and a secret semantic target in embedding space. Semantic Perinucleus Fingerprints (SPF) train the model, with SFT followed by PPO, to give a plausible but non-modal answer to a subjective question. Semantic Cluster Fingerprints (SCF) cluster the base model's own completions to a factual question and train the model, mainly with GRPO and a KL penalty to the base model, to concentrate its answers near a secret cluster centroid. Verification counts the responses whose sentence embeddings land near the target and standardizes the count against a null estimated from unfingerprinted models. On Qwen2.5-7B-Instruct, the authors report near-perfect verification scores with 1024 units under input, output, and joint paraphrase attacks, and 97.0% normalized utility for GRPO-trained SCF.
Critical strengths
- The problem is well motivated. Fingerprints that avoid exact lexical matching and use natural queries address a known weakness of memorization-based fingerprints, and supporting many independent units is practically relevant.
- The experiments vary the number of units from 16 to 1024 at a fixed verifier budget, which separates the number of units from the number of queries. SCF at 1024 units keeps a mean normalized utility of 0.970. If this survives the checks requested below, detectability under paraphrasing at near-baseline utility would be a meaningful result.
- The paraphrase attacks are designed with some care: the input rewriter is instructed to avoid the training-time paraphrases, and the output rewriter is instructed to preserve every factual claim.
- The appendix reports sensitivity analyses and failure cases, such as the collapse of some LoRA-rank and learning-rate settings and the per-concept analyses. These help readers see when the method works.
Critical weaknesses
W1. The central claim is not tested against prior methods. The abstract and Section 1 state that existing methods are either scalable (memorization-based, lexical evidence) or paraphrase-robust (statistical, watermark-style), but not both. Contribution 3 and Section 2 describe comparisons matched on owner, verifier, and attacker budgets. The experiments, however, run no prior method. The only reference points are the original model and Llama-3.1-8B-Instruct as false-positive controls, and Fig. 3 is explicitly qualitative. The premise that scalable lexical fingerprints break under paraphrasing is not measured either. Both families have public protocols: Perinucleus fingerprints (Nasery et al., 2025) and semantically conditioned watermarks (Gloaguen et al., 2025, who already evaluate input and output paraphrasing). Without at least one representative of each family at the same N and under the same attacks, the claim that this regime was previously missing is unsupported. A related ablation is also missing. SPF extends Perinucleus sampling with a 20-paraphrase training family, a PPO stage, and embedding-based verification, but the paper does not isolate what each contributes relative to token-level Perinucleus trained with the same paraphrase augmentation.
W2. Robustness is evaluated against paraphrasing and quantization (Fig. 20a, K=256) only, although the threat model includes weight access. Section 2 states that the attacker has access to the fingerprinted weights and lists robustness to suppression among the requirements, following Nasery et al. (2025). Fine-tuning, pruning, and model merging, which prior fingerprinting work routinely examines, are not evaluated, so it is unclear whether a design scaled to K=1024 survives realistic modifications. Fine-tuning is the most standard of these. Both schemes verify through a shift in the model's response distribution, so it is not obvious that the signal survives further training, and I would ask for at least one setting at K=1024, for example benign fine-tuning on a general instruction dataset.
W3. The false-positive evidence is thin, and the verification statistic is hard to interpret. The only negative controls named are the original model and Llama-3.1-8B-Instruct (Section 5.1). There are no models derived from the same base, no other models of the same family, and no model fingerprinted with a different key, and no false-positive rate over multiple negative models is reported. This matters most for SCF, whose target is the highest-scoring cluster of the base model's own completions (Alg. 8). In Fig. 1 (bottom right), the original model already reaches a target-anchor hit rate of roughly 0.2–0.4, and at K=1024 the fingerprinted model's rate is roughly 0.38–0.44 against about 0.27 for the original model (my reading of the figure), so detection rests on a modest gap and an accurate null. The null, however, is specified inconsistently. Section 3.2.3 says it is estimated from auxiliary unfingerprinted models with a fallback of 1/|A_j|, whereas Algorithm 11 uses 1/|A_k| (random anchor pick). The composition of the auxiliary set is not stated, and if the two control models belong to it, their near-zero scores in Fig. 1 are not independent evidence.
W4. "Independently assignable" units are not tested. Section 1 motivates many units by distributing different credentials to auditors, customers, or licensees and localizing leakage, and Section 2 defines a verification block as the subset of units given to one verifier. No experiment partitions the units into blocks, reports detection at the per-block budget, or checks that exposing one block leaves the others verifiable. The scaling evidence shows that z grows with K at a fixed m=20 (Fig. 11), so the total number of queries grows with K, and at the fixed budget N=1024 (Fig. 1) only saturated scores and hit rates are shown. This establishes that many units can coexist, not that they can be handed out independently. Table 9 also raises doubts about per-unit reliability. Concept-level z ranges from +267 to about −65, so some units carry little or no usable signal, and for the 14 concepts listed, the maximum cosine similarity between a concept's target anchor and another concept's anchor ranges from 0.51 to 0.93, with four values above the hit threshold τ=0.85. The per-unit success rate at K=1024 is not reported.
W5. In the main experiments, utility is preserved only by SCF, and the utility results come without any measure of uncertainty. SPF's normalized utility falls from 0.940 at K=16 to 0.661 at K=512 and returns to 0.863 at K=1024 (Fig. 2, Section 5.2). The authors suggest that this non-monotonic pattern reflects untuned hyperparameters but do not test it, and Fig. 7 uses a different model and different benchmarks, so it does not explain Fig. 2. For SCF, Fig. 2 has no error bars or confidence intervals. The baseline uses five seeds, but its spread and the number of runs for the fingerprinted models are not reported, and the curves suggest that noise is comparable to the effects being discussed (SCF's GPQA-Diamond score exceeds 1.0 for K=16–256, up to about 1.07). There are also small inconsistencies: the Introduction states "within ∼3 relative percentage points … across five benchmarks" while Section 5.1 uses four, and the caption of Fig. 13 says GRPO stays within ∼1pp of baseline on every benchmark, but the bars show GSM8K at about 0.85 against 0.88.
W6. The conclusion that RL is the appropriate objective is confounded. In SCF, the SFT baseline has no KL term or other explicit constraint to the base model, is reported only at a learning rate of 2×10⁻⁴ (Table 7), and collapses (GSM8K drops from 0.882 to 0.006, Fig. 13). GRPO, in contrast, has a KL penalty to the base model (β=0.05), which Section 3.2.2 itself credits with limiting drift. Fig. 9 shows that stopping SFT earlier does not recover utility, but that does not separate the RL objective from the regularization. Attributing the utility gap to the RL objective requires a regularized SFT baseline, or GRPO without the KL term. For SPF, SFT-only and SFT+PPO give the same belief shift of +0.30 in Fig. 16, yet the caption says SFT+PPO yields the strongest semantic shift and adds attack robustness. The figure shows no attack results.
Questions for authors
- Can you compare SCF and SPF with Perinucleus fingerprints and with Gloaguen et al. at the same N and under the same paraphrase attacks, even at a smaller K? How does token-level Perinucleus trained with the same 20 paraphrases perform?
- After benign fine-tuning at K=1024 (for example, on Alpaca-style data), what are the verification score and the utility?
- Which unfingerprinted models form the auxiliary set used to estimate p̂_0? The original model's target-anchor hit rate in Fig. 1 is about 0.2–0.4, so what score does it receive under the statistic of Alg. 11 (γ_k=1/|A_k|)? What are the scores, or false-positive rates, on additional unfingerprinted models, including models derived from Qwen2.5-7B-Instruct and a model fingerprinted with a different key?
- If the 1024 units are split into disjoint blocks (for example, of 16 units), what fraction of blocks is detected at the per-block budget? Does leaking one block affect the verification of the others?
- How many runs support each point in Fig. 2, and what are the confidence intervals? Why does the SFT+PPO mean in Fig. 7 exceed the base-model mean up to K=512?
- In Figs. 17–22, z at K=256 is between 496 and 851 for what appears to be the same chosen configuration, whereas Fig. 11 gives about 58 at K=256 and m=20. What are m, N, and the exact statistic in each figure? How large is the run-to-run variation, given that what appears to be the same configuration gives 652 in Fig. 17a but about 553 in Fig. 17b, and 530 in Fig. 19a but 496 in Fig. 19b? Table 9 reports concept-level z of up to 267. Under the binomial statistic of Alg. 11, a z of 267 at γ=0.06 requires roughly 4,500 queries for that concept, whereas Fig. 12 uses m=21. How are concept-level scores combined into the "aggregate z-score" (the caption of Fig. 10 mentions Stouffer's method)?
- What is the secret key of SCF? Alg. 8 selects the target anchor as the highest-scoring cluster of the base model's responses, whereas Fig. 5 says the owner "secretly selects" it. What can an adversary who knows the scheme and the base model recover about the targets?