- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
The paper proposes D-BISR, an encoder-based harmful prompt detector meant to transfer to unseen datasets without target data. On top of a CLIP text encoder, K=8 learnable prototypes attend over token features, and their outputs are fused with the global embedding. The fused vector is scored by a single-step denoising residual trained with three losses: a safe-only residual loss, a pairwise safe/unsafe ranking loss, and a prototype diversity loss (§4, Eq. 16, p.7). The threshold is chosen on source calibration data and applied unchanged to targets. Validation consists of in-domain and cross-dataset results on ViSU, MMA, and SneakyPrompt (Tables 1–2, p.7), a component ablation on SneakyPrompt (Table 3, p.8), and efficiency and seed-stability tables (Tables 4–5, p.9).
Critical strengths
- Clear gains on SneakyPrompt. In-domain AUROC is 0.9538 vs. 0.8846 for the best baseline (LEG), and TPR@FPR5 is 0.8397 vs. 0.2628 (CLIP-Linear) (Table 1, p.7). D-BISR also has the best TPR@FPR5 in all four transfer directions involving SneakyPrompt (Table 2, p.7). This is the main evidence for the paper’s claims, but it rests on roughly 156 unsafe test prompts and does not extend to ViSU or MMA, so it does not carry my assessment.
- A deployment-realistic protocol. Thresholds are selected on source calibration data and fixed for testing, and test splits are length-balanced (§5.1, p.7–8). This makes the low-FPR numbers meaningful, which is also why the gaps in baselines and benchmarks below matter.
Critical weaknesses
- The benchmarks, baselines, and related work do not support the paper’s claims.
- Benchmarks. The paper motivates input filtering for LLMs (§1, p.1), but all three benchmarks are text-to-image NSFW or jailbreak prompt sets (§5.1, p.7), so its LLM claims are untested.
- Missing encoder guards. Beyond three LLM guards, the only detectors discussed are HyPE and LEG (§2, p.2–3). Lightweight encoder guards with public checkpoints are missing, such as HarmAug (ICLR 2025) and Meta’s Prompt Guard and Llama Prompt Guard 2. These are the same kind of model as the paper’s own DeBERTa-v3 prompt-injection baseline.
- Missing guards for unseen attacks. Guardrails designed to adapt to new attacks without retraining, such as Retrieval-Augmented Defense (ACL 2026), are not discussed.
- Missing text-to-image guards. Given the actual text-to-image setting, Latent Guard (ECCV 2024) and GuardT2I (NeurIPS 2024) are the direct competitors, and they are neither cited nor compared.
- Baseline choice. No rationale is given for the baselines, and they differ by table. Table 1 omits DeBERTa-v3, and Table 2 omits HyPE and LEG.
- BingoGuard. BingoGuard-8B appears only in the latency table (Table 4, p.9). Yet §5.5 asserts it “achieves strong AUROC” with no number reported, so the accuracy–efficiency trade-off claim is unsupported.
- Unclear gains. HyPE has higher F1 on ViSU and MMA (0.9681 vs. 0.9592; 0.9537 vs. 0.9502; Table 1), contrary to §5.2. The MMA TPR@FPR5 margin over HyPE (0.0051) is within the seed std of 0.0118 (Table 5). In MMA→ViSU, D-BISR trails the same-encoder CLIP-L (AUROC 0.7079 vs. 0.8227; Table 2).
- The method’s contribution is neither clearly explained nor isolated.
- Readability. The paper is hard to read. Heavy notation and new labels for familiar effects obscure a simple pipeline. For example, “risk-mode entanglement” (§3.2, p.4) means that unsafe prompts form several clusters, and Eq. 2 is the k-means between-cluster variance identity.
- Figures and tables. Figures 1–2 are low-resolution raster images with illegible labels and legends (p.3–4). Table 2’s rules overlap its rows (p.7), and Tables 1 and 3–5 mark no best results. “D-BISR” is never expanded, and the dataset definition in §4 is garbled (p.5).
- Scorer not isolated. The “diffusion” scorer adds noise to one fused vector and uses the noise-prediction error at a single step as the score (§4.4–4.5, p.6–7). Removing the ranking loss drops AUROC to 0.5276 (Table 3, p.8), so the residual carries no safe/unsafe signal on its own. The “w/o diffusion” variant changes the scorer and the loss at once, so whether denoising adds anything over a ranking-trained MLP head is untested.
- Missing details. M, t, and the noise schedule are not reported, and the ablations cover only SneakyPrompt.
Questions for authors
Clarifications:
- How are safe prompts sourced, and how are safe/unsafe pairs built, for MMA and SneakyPrompt, which have no natural pairs?
- How were the baselines chosen, and why do they differ between Tables 1 and 2? Was the DeBERTa-v3 prompt-injection model fine-tuned on the source data, and what is the BingoGuard-8B AUROC referred to in §5.5?
- LEG and HyPE report a SneakyPrompt TPR@FPR5 of 0.0001 (Table 1), which is unattainable with roughly 156 unsafe test prompts. How were these values computed?
- What are M and t at inference, and how much does TPR@FPR5 vary across noise draws?
Suggestions:
- Compare against an MLP head on the same fused features trained with the same ranking loss, on all datasets with several seeds.
- Either re-scope to text-to-image and compare with Latent Guard and GuardT2I, or evaluate on LLM harmful-prompt benchmarks against HarmAug and the Prompt Guard models.
2 comments
Thank you for the detailed response. The capacity-matched MLP control is the most useful addition. At the fixed source threshold, the diffusion scorer keeps target FPR far lower at similar TPR in some directions, for example 9.6% vs. 61.5% on MMA->SneakyPrompt, which supports its role in threshold transfer. Outside MMA->ViSU, however, its AUROC matches the MLP's (0.941 vs. 0.940), so it does not improve ranking. My main concern stands. BingoGuard, never trained on these datasets, reaches AUROC 0.978-0.984, above D-BISR's in-domain AUROC on every dataset (0.954–0.975). This makes comparisons with smaller public guards such as HarmAug and Prompt Guard, and with text-to-image prompt guards such as Latent Guard and GuardT2I, essential. The response says such comparisons do not fit the source-only protocol, but none of these guards needs target data, and the authors can run them on the same target prompts as BingoGuard. The response also promises the re-scoping and presentation fixes, but the paper does not include them yet. Therefore, I maintain my score.
We thank the reviewer for the constructive feedback. Regarding benchmark relevance, D-BISR evaluates textual prompts only, without generating images or accessing visual features. Although ViSU, MMA, and SneakyPrompt originate from text-to-image safety settings, their inputs are textual prompts. We will explicitly limit our claims to text-prompt risk screening rather than general LLM or image safety. We will also discuss HarmAug, Prompt Guard, Llama Prompt Guard 2, and Retrieval-Augmented Defense, clarifying their different objectives and protocols. Their published results cannot be directly compared under our source-only cross-domain evaluation.
Regarding novelty, we do not claim new attention operators, diffusion algorithms, or ranking losses individually. Our contribution is their joint design for target-data-free detection: token-level multi-prototype attention captures localized risk cues, global–local fusion preserves context, and safety-aligned diffusion residual scoring with pairwise ranking produces scores calibrated solely on source-domain safe samples.
Regarding baselines, Table 1 evaluates within-dataset detection, while Table 2 evaluates source-to-target transfer. HyPE and LEG were excluded from Table 2 because their official implementations could not be aligned with our strict source-only training and fixed-threshold protocol. DeBERTa-v3 was initialized from a public prompt-injection checkpoint and fine-tuned for three epochs on each source domain. We will clarify these settings.
Regarding BingoGuard-8B, additional evaluation on the same target prompts yields AUROC values of 0.9784/0.9836/0.9826 on ViSU/MMA/SneakyPrompt. BingoGuard has approximately 8.03B parameters, versus 140.23M for D-BISR (57.3×). We will present it as a large-scale system reference, not a capacity-matched baseline.
Regarding performance, Table 2 reports the following macro-averages across six transfer directions:
D-BISR achieves the highest macro-averages among these encoder baselines and leads in TPR@FPR5 on four of six transfers. However, MMA→ViSU remains challenging: its original AUROC of 0.7079 trails CLIP-L (0.8227) and DeBERTa (0.8890). Across three additional full-model seeds, D-BISR obtains AUROC 0.747±0.059 and TPR 48.85±7.44%. HyPE also achieves higher within-dataset F1 on ViSU and MMA. We will explicitly acknowledge these limitations.
Regarding the diffusion scorer, we added a capacity-matched MLP control using identical fused features, safety-score constraints, and pairwise ranking supervision. Under fixed source-domain q95 calibration, D-BISR achieves lower FPR in five of six transfers and higher AUROC in four. Its macro-average FPR/TPR/AUROC is 15.41%/82.80%/0.907, versus 39.15%/85.91%/0.864 for MLP. Thus, diffusion scoring improves false-positive control in most directions, although MLP achieves higher average TPR.
Regarding inference, diffusion operates on fused text representations, not generated images or text. We use T=100, a linear beta schedule from 0.0001 to 0.02, uniformly sampled training timesteps, and fixed inference t=20, averaging residuals over M=2 independent noise samples. We will document these settings; the current results do not establish sensitivity across different M or noise samples.
Regarding paired data, ViSU provides safe/harmful fields; MMA and SneakyPrompt provide row-aligned SFW/NSFW prompts. We preserve original row indices without assuming semantic equivalence or minimal edits. Normalized-prompt connected components determine source training, selection, and calibration splits; target labels are used only for evaluation.
Regarding SneakyPrompt TPR@FPR5, the metric is the maximum empirical TPR at FPR≤5%. LEG and HyPE detected 0/156 harmful prompts, giving 0%. The reported 0.0001 was a display floor, not a measured result; we will correct it.
Finally, we will improve figure labels, legends, table formatting, notation, dataset definitions, and the explanation of D-BISR at first mention.