Search papers, labs, and topics across Lattice.
This study uncovers how deep vision models exploit invisible metadata traces at the pixel level, which correlate with supervision signals during pretraining. By introducing controlled metadata-semantics correlations, the authors demonstrate that stronger correlations lead to increased sensitivity to metadata and greater performance degradation when faced with distribution shifts. The research also identifies mitigation strategies that can reduce this sensitivity without compromising performance on downstream tasks, revealing a dual nature of metadata sensitivity in enhancing generated-image detection capabilities while improving out-of-distribution generalization.
Invisible metadata traces embedded in images can dramatically skew the performance of vision models, revealing a hidden vulnerability in their training process.
Deep vision models exploit shortcuts, relying on cues that correlate with supervision signals. Prior work has focused on visible biases, such as object-background or texture correlations. We identify a different source of shortcut learning: invisible metadata traces embedded at the pixel level, for metadata such as image processing and photo acquisition. We hypothesize that large-scale semantic supervision, whether through categorical labels (ImageNet) or billion-scale captions (LAION), naturally induces metadata-semantics correlations during pretraining, leading models to convert low-level signals into predictive features. By introducing controlled metadata-semantics correlations, we show that stronger ones produce systematically higher sensitivity to metadata traces and larger performance degradation under metadata distribution shifts. We further explore mitigation strategies applied during and after pretraining that reduce sensitivity not only to targeted metadata but also to unseen ones, without sacrificing performance on downstream tasks. Metadata sensitivity also has a positive side: it partly explains the strong generated-image detection ability of some encoders, while its mitigation can improve out-of-distribution generalization. Code: https://github.com/ryan-caesar-ramos/visual-encoder-traces
