SPOOK
A work studying the failure modes of vision encoders far from the natural image manifold. It proposes a sparse pixel hill-climbing method that optimizes on a blank canvas to probe semantic mis-mappings of encoders, revealing off-manifold fragility shared by multimodal and self-supervised encoders.
Timeline (last 90 days)
The SPOOK study uses a pixel-level hill-climbing algorithm to perform sparse pixel edits on a pure white canvas, probing vision encoders in a black-box manner.
The SPOOK experiments evaluated OpenCLIP and DINOv3 models on the ImageNet and Flickr30k datasets.
SPOOK defines the phenomenon in which a vision encoder's embedding similarity decouples from human-visible semantics when the input deviates from the natural image manifold as the "off-manifold failure mode."
A semantic basin is a region in embedding space corresponding to a specific semantic concept, and SPOOK found that sparsely editing a blank canvas can guide the encoder to map to a target semantic basin.
Owing to its black-box nature, SPOOK's method can be applied to any queryable vision encoder, including closed-source APIs.
Black-box probing methods like SPOOK are expected to become a standard diagnostic item before vision encoders are released.
All Facts (6)
The SPOOK study uses a pixel-level hill-climbing algorithm to perform sparse pixel edits on a pure white canvas, probing vision encoders in a black-box manner.
50%UnverifiedThe SPOOK experiments evaluated OpenCLIP and DINOv3 models on the ImageNet and Flickr30k datasets.
50%UnverifiedSPOOK defines the phenomenon in which a vision encoder's embedding similarity decouples from human-visible semantics when the input deviates from the natural image manifold as the "off-manifold failure mode."
50%UnverifiedA semantic basin is a region in embedding space corresponding to a specific semantic concept, and SPOOK found that sparsely editing a blank canvas can guide the encoder to map to a target semantic basin.
50%UnverifiedOwing to its black-box nature, SPOOK's method can be applied to any queryable vision encoder, including closed-source APIs.
50%UnverifiedBlack-box probing methods like SPOOK are expected to become a standard diagnostic item before vision encoders are released.
50%