Pith.

Computer Vision and Pattern Recognition · arXiv:2610.26452 v1 · submitted 19 Oct 2026 · 58 references

When Guidance Becomes a Guard: Early Interception of Unsafe Inputs in Text-to-Image Generation

Yanchun Li, Yujin Wu, Gangkai Shen, Ming Yang, Li Zeng, Xianlong Wang, Hangtao Zhang, Shujuan Tian

The paper says

“Text-to-image diffusion models can produce unsafe content, while output filtering intervenes only after synthesis. We investigate whether native generation signals enable early detection of unsafe and adversarial inputs.”

more of the abstract

We find that classifier-free guidance exposes strong safety signals in conditional responses, accessible through compact statistics and linear readouts. Based on this finding, we propose Conditional Semantic Trajectory Interception (CSTI), which aggregates activation contrasts between conditional and unconditional responses at early checkpoints for calibrated risk detection. CSTI fits concept-specific detectors using canonical concept and safe examples, without attack samples, intermediate image decoding, or generator modification. Across five generators, four adversarial attack collections, and two general NSFW datasets, CSTI achieves the strongest overall performance among six baselines at a matched 5% FPR. Early interception reduces generation latency by over 60% on intercepted paths. These results demonstrate the effectiveness of native conditional responses for early safety interception.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 20 Oct 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. A genuinely new CFG-contrast safety signal with a clean linear readout, but the layer-selection gap keeps the headline 'no attack samples' claim from being fully supported. This line is wrong

3 major objections · 5 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 58 references say.

The paper says, at Section 4.1, Section 4.3, Figure 3(a)

The review did not quote this passage.

Pith says: Objection 1 of 3 · major

The deployed observation site (U-Net middle block, SD3.5 Transformer block 11, Z-Image layer 14) is stated in Section 4.1 without any description of how it was chosen, while Section 4.3 reports full layer sweeps with large TPR swings (e.g., SDXL nudity 73.57 to 97.62; Z-Image violence 42.41 to 95.50).

more of objection 1

If those sweeps were inspected before fixing the layer for Table 2, the attack-collection labels enter the pipeline through model selection even though no attack prompt is used to fit the logistic head. Section 3.3 separates only threshold calibration from fitting; it does not protect layer, checkpoint, or statistic selection. The acknowledged exchangeability caveat in Section 3.3 and the collection-dependence caveat in Section 4.2 do not address this selection-leakage path. Please report a validation-based layer-selection protocol or pre-registered layer choices and confirm that the Table 2 configuration did not depend on the attack collections.

The paper says, at Section 3.2, Section 4.1

The review did not quote this passage.

Pith says: Objection 2 of 3 · major

The 144 canonical concept examples per concept are never specified: no prompt templates, source collections, or selection criteria are given, and no overlap audit against the six nudity and two violence evaluation collections is reported. Since the central claim is that a detector fitted without attack samples generalizes to crafted adversarial prompts, the representativeness and disjointness of these positive examples are load-bearing.

more of objection 2

Please disclose the exact construction of the canonical sets and report a duplicate/overlap analysis with the evaluation collections before the no-attack-sample generalization claim can be fully assessed.

The paper says, at Section 4.1, Metrics paragraph; Section 3.3

The review did not quote this passage.

Pith says: Objection 3 of 3 · major

The headline TPRs at a matched 5% FPR are obtained by setting the threshold on the 300 benign evaluation prompts, not by applying the safe-calibration rule of Eq. (7) at alpha = 0.075. The paper states this explicitly, but the abstract's calibrated-risk-detection claim and Table 2's numbers therefore describe a test-calibrated upper bound rather than the deployed interception operating point.

more of objection 3

Please report TPR at the alpha = 0.075 calibration operating point alongside the matched-FPR numbers; Figure 3(c) currently reports only FPR for that protocol, not the corresponding TPR.

5 minor objections

The paper says, at Section 3.3, Eq. (7)

Pith says: minor objection 1 of 5

The sentence 'The quantity p_safe is a ranking score, not the probability that the input is benign' is missing terminal punctuation; this is a typographical issue only.

The paper says, at Section 4.1, Baseline descriptions

Pith says: minor objection 2 of 5

The dagger on DAG is explained only as 'detection component only' with 'architecture-specific adaptations identified in the comparison'; please itemize these adaptations so that the baseline results are reproducible.

show 3 more

The paper says, at Section 4.4, Table 4

Pith says: minor objection 3 of 5

The sentence 'All trainable methods are adapted using the same 200 OVERT samples' should specify for CSTI what 'adapted' means (retraining the linear head, re-fitting PCA, or only recalibrating the threshold), since Section 3.2 defines a fixed fitting procedure.

The paper says, at Figure 3(a)

Pith says: minor objection 4 of 5

The layer-wise readability panel would benefit from a legend or axis labels identifying which row corresponds to which generator and concept; as printed, the layer order is not defined.

The paper says, at Abstract and Section 4.1

Pith says: minor objection 5 of 5

The abstract's count of 'four adversarial attack collections and two general NSFW datasets' does not obviously match the six nudity collections listed in Section 4.1; please clarify which collections are treated as adversarial and which as general.

What would settle it

Re-run Table 2 with the observation block fixed before seeing any attack TPR, e.g., the middle block/layer for all five generators, refitting the full pipeline (standardization, PCA, logistic head, calibration) and reporting macro TPR at 5% FPR on the same 40 settings. If macro TPR drops by more than about 2–3 points, or if IGD beats CSTI on any generator, the published Table 2 numbers are contaminated by test-set selection of the observation layer.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 20 Oct 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic partly agrees 1 min 2 s · 17,535 tokens read, 9,002 written
  3. Editor wrote the verdict above 34 s · 18,438 tokens read, 5,522 written
  4. Referee major revision: 3 major objections, 5 minor 54 s · 17,874 tokens read, 7,854 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

The thing to know: this paper taps a genuinely new signal — the conditional-minus-unconditional activation contrast inside CFG — and shows a linear readout on mean/RMS channel statistics at steps 5/10/15 gets 92–96% macro TPR at a matched 5% FPR across five generators spanning U-Net and Transformer backbones. The second thing to know: the per-generator observation layer is fixed (U-Net middle block, SD3.5 block 11, Z-Image layer 14) without a described held-out split, and layer TPR swings by 20–50 points in the same paper's sweep, so the 'no attack samples' generalization claim is not yet fully supported.

What the paper does well: the signal itself looks real — the ablations build a coherent picture, with mean+RMS beating either statistic alone, the joint checkpoint representation beating any single checkpoint, and logistic regression beating nonlinear readouts on the same features. Evaluation is careful in the ways that matter: the matched-FPR protocol is explicit, fitting is separated from threshold calibration, tables are internally consistent with the claimed macro TPRs, and the authors state the exchangeability condition for the FPR bound instead of hiding it. They also admit collection dependence and the over-refusal tradeoff.

Where it's soft: the layer-selection gap is the load-bearing issue. If the middle block or layer 14 was chosen after inspecting TPR on the attack collections, the labels enter the pipeline through model selection even though no attack prompt touches the logistic head. The calibration separation in Section 3.3 protects only the threshold. A validation split for observation-site choice, or a pre-registered rule, closes the hole. Second, the canonical concept examples (144 per concept) are neither specified nor released, and overlap with the evaluation collections is not ruled out — an addressable transparency problem. Minor: the abstract's count of 'four adversarial attack collections' doesn't match the eight collections in Table 2; that's a wording issue, not a substantive one.

Who it's for: T2I safety practitioners and anyone reading diffusion internals for discriminative tasks. It deserves a serious referee; if the authors add a proper held-out split for layer selection and release or detail the fitting sets, I'd cite it.

The strongest objection

Section 4.1 fixes the observation site per generator (U-Net middle block; SD3.5 Transformer block 11; Z-Image layer 14) without stating how the site was chosen, while Section 4.3/Figure 3(a) reports full layer sweeps with large swings (SDXL nudity 73.57→97.62; Z-Image violence 42.41→95.50). If the deployed layer was selected after inspecting TPR on the attack collections, those labels enter the pipeline through model selection even though no attack prompt is used to fit the logistic head. Section 3.3 separates only threshold calibration from fitting; it does not protect layer/checkpoint/statistic selection. The central claim, generalization without attack samples, therefore depends on an unstated assumption about how observation sites were chosen. The authors' acknowledged limitations (collection dependence in Section 4.2, over-refusal in Section 4.4, exchangeability in Section 3.3) do not address this selection-leakage path. This is an addressable methodological gap rather than a demonstrated error, so it keeps the verdict conditional.

I read the paper as claiming a specific, practical result: a detector fitted only on canonical concept and safe examples generalizes to adversarial prompt collections at a matched 5% FPR. For this to hold, no test-collection information may influence any component of the pipeline. The paper describes the logistic head and calibration as protected, but it does not describe how the observation layer was chosen. Section 4.3 shows TPR varying from 42.41 to 95.50 depending on layer, yet the chosen layers are reported without a selection procedure. If those layers were chosen after seeing attack TPRs, the main table is optimistic and the 'no attack samples' claim is formally violated through model selection. This is not an accusation of dishonesty; it is an unstated assumption in the protocol. The reader's concern about canonical-prompt provenance is related but secondary in my reading: releasing prompts would help, but would not rescue the numbers if the layer was selected on test labels. A fixed-layer rerun is a single, decisive check. Since the issue is addressable and the paper's internal logic is otherwise coherent, I keep the reader's CONDITIONAL verdict.

The referee's report

If the no-attack-sample generalization claim holds, this is a practically valuable and conceptually clean contribution: it repurposes an already-computed CFG branch instead of reconstructing images or adding decoders, uses only compact linear readouts, and is evaluated across both U-Net and Transformer backbones. Strengths include a clearly described separation between concept-head fitting and safe threshold calibration, a broad set of attack collections and baselines, and falsifiable quantitative claims whose macro averages are internally consistent. The limitations acknowledged by the authors (collection dependence, over-refusal, exchangeability of the safe calibration set) are real but do not by themselves block the claim. The blocked path is methodological transparency: the paper does not yet establish that the observation sites and canonical concept examples were selected without using the evaluation attack collections, and the headline TPRs are thresholded on the benign evaluation split rather than on the proposed calibration rule. These issues are fixable, which is why the result is worth publishing after revision.

Pith's reading of the paper (a machine's summary, not the authors')

The paper sets out to show that a text-to-image generator can be read for safety while it is still generating, rather than only after the image is complete. Its central claim is that classifier-free guidance, the standard mechanism that steers generation toward the prompt, already encodes whether the prompt is unsafe in the difference between its conditional and unconditional internal activations, and that this signal is strong within the first 15 of 50 sampling updates. The authors build CSTI, which records the per-channel mean and RMS of that contrast at updates 5, 10, and 15 and passes the concatenated vector through a whitened PCA projection and a linear classifier, with the interception threshold calibrated on benign prompts. Fitted per generator and risk concept on 144 canonical concept examples plus 256 safe prompts, with no attack samples, CSTI reports macro true-positive rates of 92.39% to 95.75% at a matched 5% false-positive rate across five generators and cuts generation latency by over 60% on intercepted paths. If correct, the result means the computation already performed to guide an image can double as its own watchdog, with no extra image decoding and no modification of the generator.

What carries the argument. The load-bearing object is the conditional contrast $D_t = H_t^p - H_t^\emptyset$, the difference between the internal activations of the prompt-conditioned and unconditioned branches of classifier-free guidance at the same sampling update. CSTI compresses each channel of $D_t$ into its mean and RMS, concatenates those $\mathbb{R}^{2C}$ vectors from updates 5, 10, and 15 into one $\mathbb{R}^{6C}$ vector, and then standardizes, whitens via PCA to 128 components, and applies a linear head whose decision value is the risk score. The interception threshold is set not by tuning on attacks but by a ranking rule over 200 independent safe calibration prompts, stopping a request when its score ranks above the calibration distribution at level $\alpha = 0.075$, which bounds the marginal false-positive rate when calibration and benign test inputs are exchangeable. The design is deliberately minimal, with a linear readout, no image reconstruction, and a frozen generator, so that the reported performance is attributable to the signal already present in the native CFG computation.

Core claim. Stated on the paper's own terms: at each sampling update, classifier-free guidance evaluates the diffusion model twice from the same latent state, once conditioned on the prompt and once on the empty input, and the difference $D_t = H_t^p - H_t^\emptyset$ between the two internal responses is a compact, early-readable safety signal. CSTI reduces $D_t$ to per-channel mean and RMS statistics at updates 5, 10, and 15, concatenates them in sampling order into a conditional semantic trajectory, and scores risk with a 128-component whitened PCA projection followed by a balanced logistic-regression head. Because the positive examples used to fit each concept detector are canonical prompts for the concept rather than successful attacks, the method claims to detect unsafe and adversarial inputs without ever seeing an attack sample, without decoding intermediate images, and without modifying the generator. The paper reports that this configuration achieves the highest true-positive rate in 37 of 40 evaluated settings against six baselines at a matched 5% FPR, that the signal is already readable at update 5, and that combining checkpoints lowers the mean empirical FPR from 15.9% at update 5 alone to 4.2% for the joint representation under its calibration protocol.

The premise it rests on. The premise the argument leans on most is that 144 canonical concept prompts per risk concept, which the paper never enumerates, produce conditional contrasts representative enough of the broad and deliberately obfuscated adversarial prompts used at test time, so that a detector fitted without any attack samples still intercepts them; the paper also acknowledges that its false-positive bound depends on safe calibration and benign test prompts being exchangeable.

If the paper is right (Pith's inferences, not the authors')

Any generator using classifier-free guidance exposes the same paired activations, and CSTI transfers across U-Net and Transformer backbones with identical statistics and readout.

Flagged requests stop after update 15 of 50, saving 61.41% to 71.72% of generation latency on intercepted paths and reducing peak GPU memory on the smaller tested models.

Because fitting uses no attack samples, a new protected concept can be covered from a handful of canonical examples plus benign prompts, before any attack against that concept is known.

Combining early checkpoints tightens false-positive behaviour: per-checkpoint empirical FPR of 15.9%, 13.5%, and 10.4% drops to 4.2% for the joint representation under the paper's calibration protocol.

The safety signal is already readable at update 5, with 88.4% nudity TPR at matched 5% FPR, so interception earlier than update 15 is consistent with the paper's measurements even though its design waits for the joint decision.

A natural extension the paper does not pursue: the conditional contrast is a semantic direction, so its geometry might not only flag unsafe prompts but also indicate which concept is being steered toward, turning interception into a triage signal for concept-specific review rather than a hard stop.

The same readout could be refitted to other protected categories, such as copyrighted style, medical misinformation, or political deepfakes, but whether the signal stays strong at update 5 for categories that are less visually saturated than nudity and violence is a testable open question.

The reported 61% to 72% latency saving is per intercepted path; a serving system's throughput gain also depends on the share of requests intercepted, a quantity governed jointly by the 5% FPR and the incoming attack rate that the paper does not model.

One stress test worth running is a semantically safe prompt phrased in an adversarial style: if it is intercepted, the method is reacting to prompt form rather than generative intent. The paper's OVERT experiments probe exactly this boundary, but only on two generators.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Text-to-image diffusion models can produce unsafe content, while output filtering intervenes only after synthesis. We investigate whether native generation signals enable early detection of unsafe and adversarial inputs. We find that classifier-free guidance exposes strong safety signals in conditional responses, accessible through compact statistics and linear readouts. Based on this finding, we propose Conditional Semantic Trajectory Interception (CSTI), which aggregates activation contrasts between conditional and unconditional responses at early checkpoints for calibrated risk detection. CSTI fits concept-specific detectors using canonical concept and safe examples, without attack samples, intermediate image decoding, or generator modification. Across five generators, four adversarial attack collections, and two general NSFW datasets, CSTI achieves the strongest overall performance among six baselines at a matched 5% FPR. Early interception reduces generation latency by over 60% on intercepted paths. These results demonstrate the effectiveness of native conditional responses for early safety interception.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-10-20

The authors' main claim, as the classifier picked it out 1.00

“These results demonstrate the effectiveness of native conditional responses for early safety interception.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.06 · states limitation 0.11 · human participants 0.03 · controlled comparison 0.92 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.97 · mentions code or data availability 0.03

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Text-to-image diffusion models can produce unsafe content, while output filtering intervenes only after synthesis. background 0.94 0.06 0.04 asserted 0.97
We investigate whether native generation signals enable early detection of unsafe and adversarial inputs. contribution 0.80 0.37 0.04 unclear 0.64
We find that classifier-free guidance exposes strong safety signals in conditional responses, accessible through compact statistics and linear readouts. result 1.00 0.96 0.37 asserted 1.00
Based on this finding, we propose Conditional Semantic Trajectory Interception (CSTI), which aggregates activation contrasts between conditional and unconditional responses at early checkpoints for calibrated risk detection. method 0.81 0.85 0.15 asserted 0.98
CSTI fits concept-specific detectors using canonical concept and safe examples, without attack samples, intermediate image decoding, or generator modification. method 0.98 0.84 0.07 asserted 1.00
Across five generators, four adversarial attack collections, and two general NSFW datasets, CSTI achieves the strongest overall performance among six baselines at a matched 5% FPR. result 1.00 0.95 0.26 asserted 1.00
Early interception reduces generation latency by over 60% on intercepted paths. result 1.00 0.93 0.09 asserted 1.00
These results demonstrate the effectiveness of native conditional responses for early safety interception. interpretation 0.99 0.87 0.80 asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.98

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-20: 1 request, 6,099 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (3)
Figure 1 of the paper
Figure 1. From late checking to early interception. (a) Crafted prompts can induce unsafe generations. (b) Post-generation safety checks incur the full generation cost. (c) In this example, CSTI reads early CFG activations at steps 5, 10, and 15, intercepting risky inputs while allowing accepted requests to continue unchanged. Intermediate images are shown only for illustration. For each generator and risk concept, the detector is fitted using canonical concept examples and safe examples with￾out attack samples. An independent safe set is used to cal￾ibrate the interception threshold,…
Figure 2 of the paper
Figure 2. Overview of CSTI. (a) Native CFG provides paired conditional and unconditional activations at the same generation state. Their difference forms a conditional contrast. (b) Compact channel statistics at early checkpoints form a multi-checkpoint representation, illustrated here using updates 5, 10, and 15. (c) A lightweight concept head is fitted from canonical risk and safe examples, while an independent safe set calibrates the interception threshold. In this example, CSTI decides once at update 15. 3.1. Early Safety Signals in Native Conditional Re￾sponses Classifier-free guidance…
Figure 3 of the paper
Figure 3. Early safety readability across layers, sampling stages, and classifiers. (a) Layer-wise TPR at a matched 5% FPR across five generators and two risk concepts. (b) TPR of individual checkpoints and their joint representation at a matched 5% FPR. (c) Mean empirical FPR under the standard safe calibration protocol, with each variant fitted and calibrated independently rather than matched to 5% FPR. (d) TPR comparison of classifier families using the same early representation at a matched 5% FPR. Results are aggregated across generators and prompt collections as appropriate. termediate…

Figures are the authors', reproduced from the paper.

References (58)
  1. [1] Anastasios N. Angelopoulos and Stephen Bates. A gentle in- troduction to conformal prediction and distribution-free un- certainty quantification.arXiv preprint arXiv:2107.07511, on PithPith's reviewarXiv
  2. [2] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In2017 ieee symposium on security and privacy (sp), pages 39–57. Ieee, 2017. 11 on Pith
  3. [3] Ziheng Cheng, Yixiao Huang, Hui Xu, Somayeh Sojoudi, Xuandong Zhao, Dawn Song, and Song Mei. OVERT: A benchmark for over-refusal evaluation on text-to-image models. InAdvances in Neural Information Processing Sys- tems, 2025. 5, 8 on Pith
  4. [4] Zhi-Yi Chin, Chieh Ming Jiang, Ching-Chun Huang, Pin- Yu Chen, and Wei-Chen Chiu. Prompting4Debugging: Red- teaming text-to-image diffusion models by finding problem- atic prompts. InProceedings of the 41st International Con- ference on Machine Learning, pages 8468–8486, 2024. 5 on Pith
  5. [5] Yoonseok Choi, Chaeyoung Oh, Hyunjun Choi, Seokin Seo, and Kee-Eung Kim. Concept removal guidance: Evidence- calibrated negative guidance for safe diffusion sampling. In Proceedings of the 43rd International Conference on Ma- chine Learning, pages 20250–20278, 2026. 2, 3, 11 on Pith
  6. [6] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rec- tified flow transformers for high-resolution image syn... on Pith
  7. [7] Rohit Gandikota, Joanna Materzynska, Jaden Fiotto- Kaufman, and David Bau. Erasing concepts from diffusion models. InProceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 2426–2436, 2023. 2, 3, 11 on Pith
  8. [8] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples.arXiv preprint arXiv:1412.6572, 2014. 11 on PithPith's reviewarXiv
  9. [9] Kai Greshake, Sahar Abdelnabi, Shailesh Mishra, Christoph Endres, Thorsten Holz, and Mario Fritz. Not what you’ve signed up for: Compromising real-world llm-integrated ap- plications with indirect prompt injection. InProceedings of the 16th ACM workshop on artificial intellige... on Pith
  10. [10] Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg. Bad- nets: Identifying vulnerabilities in the machine learning model supply chain.arXiv preprint arXiv:1708.06733, 2017. 11 on PithPith's reviewarXiv
  11. [11] Laura Hanu and Unitary team. Detoxify. GitHub repository, on Pith
  12. [12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022. 1 on PithPith's reviewarXiv
  13. [13] Jianyi Hu, Hangtao Zhang, Yi Liu, Yeqi Zeng, Li Zeng, Xianlong Wang, Rui Wang, and Leo Yu Zhang. Jevad- vbench: A benchmark and black-box attacks for reinforce- ment learning for calibrated decisions models.arXiv preprint arXiv:2609.31142, 2026. 11 on PithPith's reviewarXiv
  14. [14] Feifei Li, Mi Zhang, Yiming Sun, and Min Yang. Detect- and-Guide: Self-regulation of diffusion models for safe text- to-image generation via guideline token optimization. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, 2025. 3, 6, 11 on Pith
  15. [15] Minghui Li, Hangtao Zhang, Yanjun Zhang, Li Zeng, Chao Chen, Qiyun Shao, Wei Wan, Shengshan Hu, and Leo Yu Zhang. Fine-grained poisoning framework against feder- ated learning.IEEE Transactions on Dependable and Secure Computing, 2025. 11 on Pith
  16. [16] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maxim- ilian Nickel, and Matthew Le. Flow matching for generative modeling. InInternational Conference on Learning Repre- sentations, 2023. 5 on Pith
  17. [17] Mingrui Liu, Sixiao Zhang, and Cheng Long. Wukong framework for not safe for work detection in text-to-image systems. InProceedings of the 32nd ACM SIGKDD Con- ference on Knowledge Discovery and Data Mining, pages 891–902, 2026. 1, 3, 6, 11 on Pith
  18. [18] Runtao Liu, Ashkan Khakzar, Jindong Gu, Qifeng Chen, Philip Torr, and Fabio Pizzati. Latent Guard: A safety frame- work for text-to-image generation. InComputer Vision – ECCV 2024, pages 93–109. Springer, 2024. 2, 3, 6, 11 on Pith
  19. [19] Grace Luo, Lisa Dunlap, Dong Huk Park, Aleksander Holyn- ski, and Trevor Darrell. Diffusion hyperfeatures: Searching through time and space for semantic correspondence. InAd- vances in Neural Information Processing Systems, 2023. 3, 12 on Pith
  20. [20] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learn- ing models resistant to adversarial attacks.arXiv preprint arXiv:1706.06083, 2017. 11 on PithPith's reviewarXiv
show 38 more
  1. [21] michellejieli. NSFW text classifier. Hugging Face model repository, 2022. Accessed September 29, 2026. 2, 3, 6, 11 on Pith
  2. [22] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M ¨uller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. InInternational Con- ference on Learning Representations, 2024. 2 on Pith
  3. [23] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684–10695, 2022. 1, 2 on Pith
  4. [24] Patrick Schramowski, Manuel Brack, Bj ¨orn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappro- priate degeneration in diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22522–22531, 2023. 2, 3, 5, 11 on Pith
  5. [25] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denois- ing diffusion implicit models. InInternational Conference on Learning Representations, 2021. 5 on Pith
  6. [26] Yufei Song, Ziqi Zhou, Minghui Li, Xianlong Wang, Hang- tao Zhang, Menghao Deng, Wei Wan, Shengshan Hu, and Leo Yu Zhang. Pb-uap: Hybride universal adversarial at- tack for image segmentation. InICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Pro... on Pith
  7. [27] Yufei Song, Ziqi Zhou, Qi Lu, Hangtao Zhang, Yifan Hu, Lulu Xue, Shengshan Hu, Minghui Li, and Leo Yu Zhang. Segtrans: Transferable adversarial examples for segmenta- tion models.IEEE Transactions on Multimedia, 2026. 11 on Pith
  8. [28] Stability AI. Stable Diffusion 3.5 Medium: Model card. Hugging Face model repository, 2024. Accessed October 11, 2026. 2 on Pith
  9. [29] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. InAdvances in Neural Information Processing Systems, 2023. 3, 12 on Pith
  10. [30] Tongyi-MAI. Z-Image: Full foundation model. Hugging Face model repository, 2025. Accessed October 11, 2026. 2 on Pith
  11. [31] Yu-Lin Tsai, Chia-Yi Hsu, Chulin Xie, Chih-Hsun Lin, Jia You Chen, Bo Li, Pin-Yu Chen, Chia-Mu Yu, and Chun- Ying Huang. Ring-A-Bell! how reliable are concept removal methods for diffusion models? InInternational Conference on Learning Representations, 2024. 1, 2, 5, 11 on Pith
  12. [32] Junhui Wang, Hangtao Zhang, Zhirun Zheng, Li Zeng, Jiejun Xiao, Xi Luo, Lihua Yin, and Saiqin Long. Pvdetector: Detecting prompt injection attacks on purpose-specific llm agents through policy-violation concept analysis. InPro- ceedings of the 34th ACM International Conference... on Pith
  13. [33] Xianlong Wang, Minghui Li, Wei Liu, Hangtao Zhang, Shengshan Hu, Yechao Zhang, Ziqi Zhou, and Hai Jin. Un- learnable 3d point clouds: Class-wise transformation is all you need.Advances in Neural Information Processing Sys- tems, 37:99404–99432, 2024. 11 on Pith
  14. [34] Xianlong Wang, Hewen Pan, Hangtao Zhang, Minghui Li, Shengshan Hu, Ziqi Zhou, Lulu Xue, Aishan Liu, Yunpeng Jiang, Leo Yu Zhang, et al. Trojanrobot: Physical-world backdoor attacks against vlm-based robotic manipulation. arXiv preprint arXiv:2411.11683, 2024. 11 on PitharXiv
  15. [35] Xianlong Wang, Wenbo Pan, Shijia Zhou, Ke Li, Yuqi Wang, Zeyu Ye, Hangtao Zhang, Leo Yu Zhang, and Xiaohua Jia. Image-to-video diffusion: From foundations to open fron- tiers.arXiv preprint arXiv:2605.17248, 2026. 11 on PithPith's reviewarXiv
  16. [36] Xianlong Wang, Hangtao Zhang, Wenbo Pan, Ziqi Zhou, Changsong Jiang, Li Zeng, and Xiaohua Jia. Dual- branch robust unlearnable examples.arXiv preprint arXiv:2605.01718, 2026. 11 on PithPith's reviewarXiv
  17. [37] Yichen Wang, Yuxuan Chou, Ziqi Zhou, Hangtao Zhang, Wei Wan, Shengshan Hu, and Minghui Li. Breaking barriers in physical-world adversarial examples: Improving robust- ness and transferability via robust feature. InProceedings of the AAAI Conference on Artificial Intelligence, ... on Pith
  18. [38] Yichen Wang, Hangtao Zhang, Hewen Pan, Ziqi Zhou, Xian- long Wang, Peijin Guo, Lulu Xue, Shengshan Hu, Minghui Li, and Leo Yu Zhang. Advedm: Fine-grained adversarial at- tack against vlm-based embodied agents.Advances in Neu- ral Information Processing Systems, 38:136551–136575, on Pith
  19. [39] Zeyu Wang, Xiaodan Li, Zhiwen Li, Yuefeng Chen, and Hui Xue. InGuard: Towards generalized inner guardrail for safe text-to-image generation.arXiv preprint arXiv:2609.27620, on PithPith's reviewarXiv
  20. [40] Meng Xie, Li Zeng, Hangtao Zhang, Xianlong Wang, Ziqi Zhou, Pengpeng Qiao, and Zhetao Li. Typo: Instruction- dense visual jailbreaks against commercial closed-source image-generation models.arXiv preprint arXiv:2607.24897, on PithPith's reviewarXiv
  21. [41] Fan Yang, Yihao Huang, Jiayi Zhu, Ling Shi, Geguang Pu, Jin Song Dong, and Kailong Wang. Seeing it before it happens: In-generation NSFW detection for diffusion-based text-to-image models.arXiv preprint arXiv:2508.03006, on PithPith's reviewarXiv
  22. [42] Jinghan Yang, Yihe Fan, Xudong Pan, and Min Yang. Flow- Guard: Towards lightweight in-generation safety detection for diffusion models via linear latent decoding.arXiv preprint arXiv:2604.07879, 2026. 1, 3, 11 on PithPith's reviewarXiv
  23. [43] Yijun Yang, Ruiyuan Gao, Xiaosen Wang, Tsung-Yi Ho, Nan Xu, and Qiang Xu. MMA-Diffusion: MultiModal attack on diffusion models. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 7737–7746, 2024. 1, 2, 5, 11 on Pith
  24. [44] Yijun Yang, Ruiyuan Gao, Xiao Yang, Jianyuan Zhong, and Qiang Xu. GuardT2I: Defending text-to-image models from adversarial prompts. InAdvances in Neural Information Pro- cessing Systems, 2024. 2, 6, 11 on Pith
  25. [45] Zeming Yao, Hangtao Zhang, Yicheng Guo, Xin Tian, Wei Peng, Yi Zou, Leo Yu Zhang, and Chao Chen. Reverse back- door distillation: Towards online backdoor attack detection for deep neural network models.IEEE Transactions on De- pendable and Secure Computing, 21(6):5098–5111, 2024. 11 on Pith
  26. [46] Zeyu Ye, Yanchun Li, Sibei He, Meng Xie, Hangtao Zhang, Xianlong Wang, Li Zeng, Jiahao Chen, Yichen Wang, Junhui Wang, et al. False claims, credible images: A red-teaming benchmark for commercial image generators.arXiv preprint arXiv:2610.11112, 2026. 11 on PithPith's reviewarXiv
  27. [47] Z-Image Team. Z-Image: An efficient image generation foundation model with single-stream diffusion transformer. arXiv preprint arXiv:2511.22699, 2025. 2 on PithPith's reviewarXiv
  28. [48] Li Zeng, Xiaojun Mo, Meng Xie, Hangtao Zhang, Yixiang Liu, Yezhuo Peng, and Yanchun Li. Psfd: Proactive spatial- frequency defense against malicious exemplar-guided image editing. In2025 IEEE International Conference on Multime- dia and Expo (ICME), pages 1–6. IEEE, 2025. 11 on Pith
  29. [49] Li Zeng, Zeyu Ye, Meng Xie, Hangtao Zhang, Xianlong Wang, Yanchun Li, and Zhetao Li. Ghostprompt: Cross- image adversarial prompt for vision-language models. In Proceedings of the 34th ACM International Conference on Multimedia, 2026. 11 on Pith
  30. [50] Hangtao Zhang, Zeming Yao, Leo Yu Zhang, Shengshan Hu, Chao Chen, Alan Liew, and Zhetao Li. Denial-of-service or fine-grained control: Towards flexible model poisoning at- tacks on federated learning. InProceedings of the Thirty- Second International Joint Conference on Artifi... on Pith
  31. [51] Hangtao Zhang, Shengshan Hu, Yichen Wang, Leo Yu Zhang, Ziqi Zhou, Xianlong Wang, Yanjun Zhang, and Chao Chen. Detector collapse: Backdooring object detection to catastrophic overload or blindness in the physical world. In 10 Proceedings of the Thirty-Third International Joint... on Pith
  32. [52] Hangtao Zhang, Yichen Wang, Shihui Yan, Chenyu Zhu, Ziqi Zhou, Linshan Hou, Shengshan Hu, Minghui Li, Yan- jun Zhang, and Leo Yu Zhang. Test-time backdoor detection for object detection models. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 2437... on Pith
  33. [53] Hangtao Zhang, Chenyu Zhu, Xianlong Wang, Ziqi Zhou, Changgan Yin, Minghui Li, Lulu Xue, Yichen Wang, Sheng- shan Hu, Aishan Liu, et al. Badrobot: Jailbreaking embod- ied llm agents in the physical world. InProceedings of the Thirteenth International Conference on Learning Rep... on Pith
  34. [54] Hangtao Zhang, Yucheng Zhao, Sishun Liu, Ziqi Zhou, Zeyu Ye, Wei Wan, Minghui Li, Shengshan Hu, Yanjun Zhang, Yi Liu, and Leo Yu Zhang. Defending jailbreak attacks on large language models via manifold trajectory kinetics. In35th USENIX Security Symposium (USENIX Security 26),... on Pith
  35. [55] Yimeng Zhang, Jinghan Jia, Xin Chen, Aochuan Chen, Yi- hua Zhang, Jiancheng Liu, Ke Ding, and Sijia Liu. To gen- erate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images ... for now. InEuropean Conference on Computer Vision, 2024. 5 on Pith
  36. [56] Ziqi Zhou, Menghao Deng, Yufei Song, Hangtao Zhang, Wei Wan, Shengshan Hu, Minghui Li, Leo Yu Zhang, and Dezhong Yao. Darkhash: A data-free backdoor attack against deep hashing.IEEE Transactions on Information Forensics and Security, 2025. 11 on Pith
  37. [57] Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J Zico Kolter, and Matt Fredrikson. Universal and transferable ad- versarial attacks on aligned language models.arXiv preprint arXiv:2307.15043, 2023. 11 A. Broader Context and Related Work Adversarial robustness and AI secur... on PithPith's reviewarXiv
  38. [2020] Accessed September 29, 2026. 2, 3, 11 on Pith

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment