- Rating
- 2
- Confidence
- 2
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
A vision-language model is put through safety unlearning so it refuses certain harmful questions. The authors ask whether a small change to the image brings the answer back (undo the unlearning!).
- The attacker is assumed to hold the model's weights and the harmful answer that unlearning was meant to suppress (whitebox method).
- For each item, ReSight runs 500 sign-gradient steps that steers every pixel of the clean image, within 32/255, so the model becomes more likely to produce the first 16-32 tokens of the answer. The perturbed image is then re-queried with the original question and no reference, and GPT-5.5 judges whether the free response matches the answer we expect to see.
- An attached qualifies as success only if 3 controls come back negative: (1) the clean image, (2) a different image carrying the same perturbation, and (3) a white image carrying it.
- Experiments cover
LLaVA-1.5-7BandQwen2.5-VL-3B, threeSafeEraserrecipes (PO, POPD, NPO+PD). - The paper reports success on ~4 to 33% of items. The highest qualified rate in 11 of 12 settings against five adapted visual attacks. POPD is the most vulnerable recipe.
Critical strengths
- POPD is the most vulnerable recipe is the most vulnerable recipe in all four model/dataset pairs.
- Detailed reporting and experiments
Critical weaknesses
- Practical setup and use case: I could not identify a setting in which this evaluation would be used. The threat model gives the attacker the model's weights, its gradients, and the harmful answer itself. I am not sure how useful or realistic this setup is.
- Findings are largely expected; Adversarial images are already known to bypass VLM safety behavior [1-3], and optimizing an image against teacher-forced target likelihood is established. ReSight is standard targeted sign-PGD on a reference-response prefix. These papers study safety-aligned models rather than these unlearning recipes, so the experiment is not redundant. But successful elicitation alone adds little.
References:
- 1 - Qi, Xiangyu, et al. "Visual adversarial examples jailbreak aligned large language models." Proceedings of the AAAI conference on artificial intelligence. Vol. 38. No. 19. 2024.
- 2 - Bailey, Luke, et al. "Image hijacks: Adversarial images can control generative models at runtime." arXiv preprint arXiv:2309.00236 (2023).
- 3 - Shayegani, Erfan, Yue Dong, and Nael Abu-Ghazaleh. "Jailbreak in pieces: Compositional adversarial attacks on multi-modal language models." International conference on learning representations. Vol. 2024. 2024.
Questions for authors
-
How much of the target response actually depends on information in the image? Could the Parent answer these questions without an image, and can independently optimized neutral or unrelated images elicit the same behavior? The current transferred-perturbation controls establish input dependence, but do not establish that meaningful visual content is necessary.
-
Contribution and practical relevance: What is the main new insight about unlearning beyond existing visual jailbreaks, and what should someone building or releasing an unlearned model do differently because of it? In particular, how do the joint controls change the assessment of an unlearning method beyond changing its reported success rate?
-
The white base is 255 in every channel, so clipping after adding the perturbation removes all positive components. Why not use a mid-grey image at 128, which would preserve the entire perturbation under the 32-level budget? How would this affect null-control matches and qualified success rates?
4 comments
Response to Reviewer 98EZ (Part 2/2: Q2, Q3)
Response 2/2: Q2 and Q3. Notation: SE/BV = SafeEraser-240/BeaverTails-V-108, L/Q = LLaVA/Qwen, V/P = Victim/Parent counts, pp = percentage points, Full = attack+ with clean/foil/white−, and Full+grey adds RGB-128 null−.
Q 2: What is the main new insight, what should a developer do differently, and what do joint controls change beyond the success rate?
Our main finding is that clean-input refusal does not guarantee local neighborhood suppression, so developers should prefer NPO+PD (4.2% to 13.9% reactivated) over POPD (12.9% to 33.3%, despite higher prefix CE) and separate three gated audit regimes:
Clean+: Clean unlearning itself failed, requiring stronger forget-set coverage or weight updates before robustness testing.Clean−, Attack+, Foil/Null+: The perturbation transfers to unrelated or blank images, indicating a generic base-VLM jailbreak rather than forget-item retention.Clean−, Attack+, All Controls−(Full): Base-image-specific local reactivation, indicating thatPOPDleaves substantially higher neighborhood vulnerability thanNPO+PD.
Joint controls sharpen near-tied fixed-image replay gaps (Table 3, SE/L +0.8 to +8.3 pp) and reverse raw checkpoint rankings under symmetric Parent/Victim PGD (n=12/cell, 500 steps on each model's own gradients):
| Gate, Δ(V−P) in pp | SE/L | SE/Q | BV/L | BV/Q |
|---|---|---|---|---|
| A. Manuscript POPD cohort (n=240/240/108/108, GPT-5.5) | ||||
| Full | +8.3 | +9.6 | +24.1 | +21.3 |
| No clean / No foil / No null | +4.6 / +7.5 / +7.9 | +9.6 / +10.4 / +9.6 | +21.3 / +24.1 / +22.2 | +21.3 / +20.4 / +19.4 |
| Attacked only | +0.8 | +9.6 | +9.3 | +13.9 |
| B. Symmetric independent PGD (n=12/cell, GPT-6.1-sol) | ||||
| Full (= Full + grey, V/P) | +16.7 (3/1) | +8.3 (2/1) | +25.0 (4/1) | +25.0 (4/1) |
| No clean / No foil / No null | +8.3 / +16.7 / +16.7 | +8.3 / +16.7 / +8.3 | +16.7 / +25.0 / +25.0 | +8.3 / +25.0 / +25.0 |
Control specificity (Full / Attacked) | V 3/4 vs P 1/5 | V 2/3 vs P 1/1 | V 4/4 vs P 1/6 | V 4/4 vs P 1/4 |
In fixed-image replay (Group A), Δ(V−P) is positive throughout, and controls expose a +8.3 pp SE/L gap (41 vs. 21) masked by raw Attacked only (41 vs. 39, +0.8 pp). Sign reversal arises in symmetric PGD (Group B), where raw Attacked counts make Victim look safer than Parent on SE/L (4/12 vs. 5/12, −8.3 pp) and BV/L (4/12 vs. 6/12, −16.7 pp), whereas joint controls reverse both symmetric-PGD rankings to +16.7 and +25.0 pp by keeping 13/15 (86.7%) Victim hits and rejecting 12/16 (75.0%) Parent clean or transfer confounders (13/48 vs. 4/48 overall, +5.0, +9.0, +19.4, +16.1 pp on frozen both-clean−, and 12/48 vs. 5/48 [+14.6 pp] across 48 prospective both-clean− PGD pairs). Joint controls separate pair-specific reactivation from clean persistence and generic transfer, preventing reversed symmetric-PGD rankings and favoring NPO+PD over POPD.
Q 3: White clips positive perturbation components. What happens with RGB-128 grey?
Adding or substituting unclipped mid-grey (RGB-128) for white (RGB-255) across all 696 POPD pairs retains 137/139 (98.6%) of primary GPT-5.5 events and 100% of unified-judge events across all radii (8, 16, 24, 32/255). As you noted, positive components carry 49.1% to 50.5% of perturbation mass and clip at 255, whereas around RGB 128 every integer stays within with zero clipping (verified on 696/696 = 100% of pairs):
| Null-control comparison (n=240/240/108/108) | SE/L | SE/Q | BV/L | BV/Q |
|---|---|---|---|---|
| A. Frozen primary GPT-5.5 + new RGB-128 grey | ||||
| Published Full (white null) | 41 | 31 | 31 | 36 |
| Full + RGB-128 grey (98.6% retained) | 41 | 31 | 29 | 36 |
| RGB-128 grey replaces white | 41 | 31 | 30 | 36 |
| B. Common blinded GPT-6.1-sol across radii | ||||
| 32/255: Null-blocked (white = grey) | 239 = 239 | 240 = 240 | 100 = 100 | 107 = 107 |
| 32/255: Full (white = grey = white+grey) | 62 | 36 | 43 | 55 |
| 24/255: Full (white = grey = white+grey) | 45 | 26 | 37 | 43 |
| 16/255: Full (white = grey = white+grey) | 10 | 9 | 14 | 12 |
| 8/255: Verified Full (= Full+grey) | 5/240 (contr., 1/6 PGD) | 2/240 (contr., 1/6 PGD) | 3/108 (contr., 4/6 PGD) | 6/108 (contr., 1/6 PGD) |
Under unified judging (Group B), requiring white, grey, or both produces the exact same qualified event set across all 16 cohort × radius cells and all 12 independent PGD cells (including all 13/13 symmetric Victim events in Q2). We will include RGB-128 grey alongside white as a standard null control in the revision. Unclipped RGB-128 mid-grey removes positive-component clipping while preserving 98.6% of primary events and 100% of unified-judge events from 8/255 to 32/255, ruling out white-clipping inflation.
Thank you again for your constructive feedback. We welcome any further discussion! : )
Thank you for the detailed response.
The grey image results answer my third question. Thanks for running them.
On the image question: you show that a grey or random image works about as well as the original image. To me, this means the image content is not what brings the answer back. The answer seems to come from the language model, and the image is just a way to push it there. The null check shows that the noise only works on its own image. It does not show that the answer needs the image.
On the use case: I understand the idea of a developer testing the model before release. But if the model is released with open weights, anyone can fine-tune it or remove the refusal directly, which is much easier than this attack. So passing this test does not make the release safer. Also, the attack stops working after a one-pixel shift, so it is not a realistic threat after deployment.
I will keep my score.
Thank you for reading our response and confirming that the mid-grey runs resolved Q3. We address your two remaining points below.
1. On the image question (answer in the LLM, image as the push).
Your reading of the grey and random image results is completely right: on these benchmarks, the target answer still sits in the language model, and the perturbed image is a continuous pathway pushing the LLM past refusal. Because a frozen CLIP encoder only outputs patch vectors and cannot write a 96-token guide on its own, the recovered text comes from the LLM decoder. We also agree that our null check proves a noise pattern is tied to its own base image, rather than proving the answer requires the original scene, which we will state plainly in Section 3.3.
Why does auditing around the original image with clean, foil, and null controls still give a critical new insight? First, this mechanism is the exact blind spot of multimodal unlearning [4, 5]: after POPD unlearning, the clean photo refuses (~98%) and asking with no image fails (22/24 = 91.7%), leading clean and text-only tests to conclude the LLM forgot, when continuous visual vectors still unlock the answer in the frozen LLM [1, 2, 4] (while NPO+PD cuts that recovery by 2.2x to 6.2x, 4.2% to 13.9% vs. 12.9% to 33.3%, despite lower prefix loss). Second, multimodal unlearning methods [5, 6] train on specific (photo, question) pairs, such as asking how to breach a lock or camera in the photo (Figures 17 to 19), and claim forgetting because that photo refuses. Perturbing that exact photo (down to 8/255) shows that uploading almost the identical photo still triggers the answer. Third, as you noted in W2, optimizing an image without controls just rediscovers generic VLM jailbreaks [1-3] that hit Parent too: in Q2, raw PGD hits Parent more than Victim (5/12 vs. 4/12 SE/L, 6/12 vs. 4/12 BV/L) due to clean non-refusals and generic triggers. Clean, foil, and null checks filter those out and flip the ranking (+16.7 and +25.0 pp), isolating post-unlearning fragility.
2. On the practical use case (API deployments, open weights, and one-pixel shifts).
Frozen-weight visual auditing serves two real settings, and one-pixel shifting does not remove the risk:
First, hosted VLM APIs and enterprise services: multimodal unlearning is built for web APIs removing private records, copyrighted content, or unsafe guides before API serving [4, 5, 6, 7]. External API users never hold server weights to fine-tune and can only upload images and text. The provider holds weights to run white-box tests before launch (and as APIs often unlearn open backbones like LLaVA or Qwen2.5-VL, surrogate transfer attacks work without server weights [2]). Catching POPD's 33.3% leakage vs. NPO+PD's 4.2% to 13.9% makes the API safer when clean tests rate both at ~98% and prior attacks [1-3] catch 0.0% to 5.6%.
Second, open-weight compliance and diagnosis: while an attacker with raw weights can fine-tune them [8, 9], open checkpoints are also deployed as-is by benign developers for privacy or safety compliance [6, 7], where leaking 31.8% of forgotten answers on unmodified weights is a direct failure. Also, fine-tuning mixes surviving memory with gradient re-teaching [9] and breaks nearly any open model, whereas a frozen-weight audit [4, 8] is a necessary first gate separating shallow refusal (POPD, 31.8% strict recovery) from stronger suppression (NPO+PD, 12.1%).
Third, on the one-pixel shift: default LLaVA and Qwen2.5-VL processors are deterministic with zero pixel shift, so saved PNGs work on standard deployments (192/192 runs). Section 5.4 shifted static PNGs without re-optimizing. By Athalye et al. [10], single-grid PGD aligns to the grid, while adaptive attackers optimize over shifts (EoT, Section 6). Above all, shifting pixels never deletes a forgotten answer from weights: one valid PNG eliciting a 96-token forbidden guide proves it was not erased.
We will make this mechanism and these use cases clear in Sections 1 and 3.3, and we sincerely hope you might reconsider your score! : )
References (continuing [1-3] in your review)
[4] Patil et al., "Unlearning Sensitive Information in Multimodal LLMs: Benchmark and Attack-Defense Evaluation," TMLR, 2024.
[5] Chen et al., "SafeEraser: Enhancing Safety in Multimodal Large Language Models through Multimodal Machine Unlearning," Findings of ACL, 2025.
[6] Liu et al., "Protecting Privacy in Multimodal Large Language Models with MLLMU-Bench," NAACL, 2025.
[7] Shi et al., "MUSE: Machine Unlearning Six-Way Evaluation for Language Models," ICLR, 2025.
[8] Łucki et al., "An Adversarial Perspective on Machine Unlearning for AI Safety," TMLR, 2025.
[9] Hu et al., "Unlearning or Obfuscating? Jogging the Memory of Unlearned LLMs via Benign Relearning," ICLR, 2025.
[10] Athalye, Carlini, and Wagner, "Obfuscated Gradients Give a False Sense of Security: Circumventing Defenses to Adversarial Examples," ICML, 2018.
Response to Reviewer 98EZ (Part 1/2: W1, W2, Q1)
Thank you for reviewing ReSight's audit setup, novelty, and controls. Part 1: W1, W2, Q1 (Q2, Q3 in Part 2). SE/BV = SafeEraser-240/BeaverTails-V-108, L/Q = LLaVA-1.5-7B/Qwen2.5-VL-3B, Parent/Victim = pre-/post-unlearning, Full = attack+ & clean/foil/white−.
ReSight is a pre-release white-box audit run by the model developer before open-weight release, rather than an external black-box attack. In open-weight unlearning benchmarks (TOFU, Maini et al., 2024, WMDP, Li et al., 2024, SafeEraser), the team releasing an unlearned checkpoint already holds weights and forget targets. The reference prefix is used only during search. Evaluation feeds only the saved 8-bit PNG and original question (zero target tokens, zero weight updates) under 96-token greedy decoding. ReSight provides a practical pre-release check for open-weight VLM developers, where auditor access is native and any failure yields a standalone, hint-free PNG exploit at deployment.
We keep sign-PGD on saved 8-bit PNGs minimal so elicited outputs reflect checkpoint behavior rather than optimizer tricks. Our contribution is formulating VLM unlearning verification as a paired-checkpoint counterfactual neighborhood audit, revealing three findings beyond single-model PGD:
0.0% to 5.6%vs. ReSight4.2% to 33.3%): Existing attacks (Qi et al., AAAI 2024 [1],Bailey et al., ICML 2024 [2],Shayegani et al., ICLR 2024 [3]) evaluate single models without paired checkpoints or transfer gates. While Table 8 spans differing threat models rather than equalized budgets, Image Hijacks [2] (ε=32/255,500steps, shared universal trigger) scores 0.0% in 7/12 settings (0.0% to 5.6%) and Jailbreak in Pieces [3] (5,000per-item CLIP steps atε=16/255) scores0.0% to 4.6%, whereas ReSight retains3.8% to 13.0%cohortFull(16.7% to 66.7%fresh PGD) atε=16/255.Parent clean+ -> Victim clean− -> ReSight+items with negative foil/null outputs (passing controls), POPD recovery reaches 10/40 (25.0%), 3/19 (15.8%), 13/32 (40.6%), and 9/19 (47.4%) (35/110 = 31.8%POPD,70/312 = 22.4%across all 12 PO/POPD/NPO+PD cohorts).0.0146 vs. 0.0371SE/Q,0.1415 vs. 0.1817BV/L,0.0574 vs. 0.2217BV/Q) yet 2.2× to 6.2× lower 96-token reactivation (4.2% to 13.9%vs.12.9% to 33.3%), showing that fitting initial prefix tokens does not equate to free-generation recovery across recipes.Paired-checkpoint counterfactual auditing shows that standard universal and CLIP-space jailbreaks [1-3] miss local retention (
0.0% to 5.6%) and that low prefix CE is not equivalent to qualified free-generation recovery.Meaningful original-image content is not universally necessary under fresh white-box optimization (
12/24blank-grey and10/24unrelated-image PGD runs succeed), although clean Parent answers require an image (22/24 = 91.7%text-only failure) and each optimized δ⋆ depends on its tested base canvas (686/696 = 98.6%blocked on grey):no-image−, n=6)Atk+ / Full, n=12)On pre-fixed POPD subsets (1 item/category), fresh PGD on neutral grey (
12/24 = 50.0%) and unrelated images (10/24 = 41.7%) shows that visual-token perturbations can elicit many target behaviors in the decoder without original-image semantics. Conversely, our null gate S(fu(g+δ⋆,q))=0 (98.6% blocked across cohorts, 15/15 = 100% in row 3) proves that δ⋆ optimized on x is specific to base image x rather than a universal trigger. Original-image semantics are not universally required when re-optimizing from scratch (12/24grey,10/24unrelated), whereas our null gate establishes base-image dependence of δ⋆ (98.6%non-transfer), which we will state explicitly in Section 3.3.Continued in Part 2.