- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3
Read review
Paper summary
The paper investigates whether preprocessing steps that are often applied to images before they are passed on to a Vision Language Model (VLM) may prevent the information that the model needs to make certain decisions from reaching the model. The authors find that the downsampling employed in such pipelines uses fixed-point integer weights, this can be exploited by constructing a pattern which yields the same results after the preprocessing but which changes the label that should be assigned to the actual image (they describe this as adding a perturbation which is in the "null space" of the downsampling) and call such a pair of images a "certified pair". Testing or verifying the correctness of the downstream model will never identify the issues introduced before the preprocessing step. An algorithm called AliasForge for constructing such perturbations is presented by the authors and is demonstrated on Qwen2.5-VL-7B, LLaVA-1.5-7B and VisualPRM-8B as well as other processors. The authors further prove that every fixed-point downscaling algorithm has such a null space and bound their smallest size. They also study whether verifiers (here this term is used to describe a model that judges whether in a reasoning mechanism a prediction made by a model is correct) and routers (which take the verdict and make a decision based on it - rerun the verifier, look at a subsection of the image in more detail, ...) can be wasteful given that they only look at postprocessed images.
Critical strengths
- The authors introduce a thorough approach for checking whether their reimplementation of a resize step is identical to that of a library and openly state when no decision can be made. The main proofs also look correct to me and the main issue that the paper targets seems like one worth investigating.
- The experimental design is thorough, verdicts were calculated before any certified pairs were computed. The authors employ various controls to double-check the different pieces of the paper, and it is nice to see that the vulnerability identified in the paper also holds up in popular models that are deployed in practice such as CLIP, SigLIP or LLaVa.
Critical weaknesses
- The main movation of the paper is that it is difficult to decide whether a model failing for a certain task is caused by a lack of visual information in the input or by the model itself. However, the authors find and acknowledge that their certified pairs can't answer this question. The pairs are stripe patterns which only hold up for specific downsampling settings, so models processing images at the original resolution or resize them using floats are not affected by this. The experimental evaluation then checks whether routers waste compute, but a router's performance on certified pairs in Table 11 seems to mostly be unrelated to its performance on real inputs. Moreover, the cost parameter is assumed to be 0 in a large part of the experimental evaluation, so I am unsure whether the "wasted compute" in Table 2 isn't mostly a consequence of that assumption. Overall, these points seem to limit the practical relevance of the paper.
- I found the router study difficult to interpret. I am not familiar with the literature in this area, but I found some things here unclear. The IAS score that the authors propose only see an image and not a question or the model's reasoning, so it doesn't seem surprising to me that from only the image information a simple classifier would not be able to decide whether to spend more compute. I also found it unclear what exactly the baseline is here, the authors introduce a "placebo" where the verifier gets the same critique prompt without being told to look at the image again, the appendix shows that this is ~7 points worse than the verifier itself. Against the original veridct, actions are 7-9 points worse and even picking the best crop is slightly worse. I'm not sure if the "stop" baseline mentioned in the experiments section is the placebo or the original verdict, but when compared to the original verdict it seems that every action makes the verifier perform worse. This performance is explained with an upper bound on what a router can gain from the processed image, but the routers are only trained on 1280 items with thousands of inputs, so the routers might just not train properly on the small training set. I don't think that this can be ruled out.
- The paper is full of jargon, new terms that (to me at least) seem to be defined by the authors somewhere and then referenced again so the reader has to keep scrolling around the paper to check again what these terms meant. I also found significant portions of the paper very difficult to understand, even when re-reading sentences multiple times. The writing should be substantially improved.
Questions for authors
- Could the authors clarify if "stop" is the placebo or the original verdict?
- The idea seems related to invariance-based adversarial examples ("Excessive Invariance Causes Adversarial Vulnerability", Jacobsen et al., 2018 and "Fundamental Tradeoffs between Invariance and Sensitivity to Adversarial Perturbations", Tramel et al., 2020), could the authors discuss the connection between those papers and their work?