Vision encoders can misinterpret sparse edits on blank images
SPOOK: Sparse Pixel Optimization for Off-Manifold Knowledge in Vision Encoders
Computer Vision and Pattern RecognitionComputation and LanguageHuman-Computer Interaction
Summary
Modern AI systems that understand images often rely on recognizing patterns similar to real photos. This paper shows that even small changes, like editing a few pixels on a blank image, can trick vision encoders into thinking the image represents specific objects or concepts. The researchers found these odd cases occur in different types of vision models, revealing a surprising blind spot in how these systems understand images. Their method creates a new way to study and better understand the shapes of visual representations AI systems learn.
What this means in practice
- •For machine learning engineers: Test whether vision models are vulnerable to sparse pixel edits that cause incorrect semantic interpretation, improving model robustness evaluation.
- •For security analysts: Detect potential exploits or failures in vision-based systems caused by minimal pixel modifications on blank or unusual inputs.
Authors
Ankit Gupta, Akshat Pandya, Christoph Adami, Emily Dolson
Abstract
Modern vision encoders are widely used as semantic representations, but the relationship between embedding similarity and human-visible semantics can become unclear far from the natural-image manifold. We probe this regime in a black-box setting by applying sparse pixel edits to blank canvases with a simple pixel-level hill-climbing procedure. Across ImageNet and Flickr30k, we evaluate whether these edited blank canvases can retrieve chosen targets in multimodal contrastive encoders and self-supervised visual encoders. On OpenCLIP, inputs optimized toward class prompts achieve strong top-1 and top-5 targeted retrieval performance; on DINOv3, inputs optimized toward class prototypes achieve high target alignment and near-top prototype ranks. These results show that sparse edits on otherwise blank, visually non-semantic inputs can enter target semantic basins in both language-aligned and self-supervised visual representation spaces. Our findings identify a shared off-manifold failure mode of modern vision encoders and suggest sparse blank-canvas optimization as a simple probe of representation geometry.