- Rating
- 2(+1)
- Confidence
- 3(−1)
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper examines whether safety information in VLM hidden states is linearly separable, and argues that separability alone does not mean those directions can be driven from the image. The evidence comes from layer sweeps on three open-weight models. Probe accuracy stays near ceiling in the late layers, yet attack success peaks much earlier, at a normalized depth of roughly 0.4–0.5. Building on this gap, the authors propose an attack called IBT. It first fits a linear safety probe at a given layer. The clean representation is then reflected across the probe's decision boundary, and the reflected point becomes the optimization target. A bounded image perturbation is optimized toward this target with an alignment loss, supplemented by a cosine term and a pixel penalty. The perturbed image is then paired with a benign-looking rewrite of the prompt, which GPT-4 generates and which is cached in advance.
Critical strengths
The paper reports a clear and practically useful empirical finding. The most linearly separable safety representation is not necessarily the best supervision target for input-space attacks. This holds across three quite different VLM families. Both the controlled layer sweep and the full IBT evaluation point to the same fairly narrow intermediate region, around normalized depth 0.4–0.5, even though late-layer probes remain highly accurate. What matters more is that the pattern does not rest on a single cherry-picked layer. A shared normalized position near 0.45 keeps almost all of each model's peak performance, with mean ASR and SASR coming very close to the independently selected optima. The finding is useful because it turns supervision-layer selection into an explicit design choice for representation-guided attacks, instead of letting the final layer, or the most separable one, serve as the default.
Critical weaknesses
The experiments show a specific phenomenon that supervision around depth 0.45 is substantially better than supervision at later layers, but the theory provides no reason for such an intermediate optimum to exist. In fact, the appendix explicitly shows that the full first-order input-to-behavior gradient is invariant to the location of the computational cut. The proposed decomposition can explain why probe accuracy alone is insufficient, but it cannot explain why L12 should outperform L27, or why three models should peak around the same normalized depth. Thus, the central theoretical story stops exactly where the paper’s main empirical claim begins.
The second concern is that the reported gains are not cleanly attributable to IBT itself. The attack uses EOT with JPEG compression, random resizing, and average pooling during optimization, but this component is never removed or separately evaluated. This is a major confound because exactly this kind of transformation-robust optimization is known to help adversarial examples survive preprocessing, transfer across models, and remain visually stable. Yet the paper attributes the strong black-box transfer and stealth results mainly to intermediate-layer targeting. The component study does not resolve this issue because EOT remains bundled inside the image optimization term.
Questions for authors
Qwen is swept only from L11 and Llama/GLM from L14, while all reported optima occur very early in the evaluated range. Can you extend the sweep substantially below d=.35?
In Table 6, why is image-only SASR below the untouched baseline, and why does the implied NIQE weight at Qwen L12 differ so much from IBT at the same layer and budget?
After looking at the paper again, I misread how SASR is computed in my second question, and I will raise my score from 1 to 2 for this paper's strength part as I wrote with the author's clarification on my two questions. The remaining weaknesses still need to be addressed.