Event
SAE Feature Suppression Defense Against VLA Adversarial Patches
1 reports · 1 sources
Yukiya Horiba and co-authors propose a defense method against adversarial patches targeting vision-language-action (VLA) models: using a sparse autoencoder (SAE) to analyze the model's internal representations, they identify a feature whose activation is strongly correlated with adversarial patches, and at inference time suppress that feature only when a linear probe detects an attack, without fine-tuning the VLA model. In adversarial patch attack tests on LIBERO-10, this conditional intervention improved the success rate under intermittent attacks; however, the authors also report that continuously applying the same intervention markedly degrades policy performance. In other words, the method's effectiveness depends on the conditional trigger of "detect first, then suppress," rather than unconditionally disabling the feature.
Report timeline
- Detect and Suppress: A Mechanistic Defense Against Adversarial Patches on VLA Models
The authors use a sparse autoencoder (SAE) to analyze the internal representations of a VLA model, identify a feature whose activation strongly correlates with the presence of adversarial patches, and suppress that feature at inference time only when a linear probe detects an attack, thereby improving robustness without fine-tuning the VLA. In adversarial patch attack tests on LIBERO-10, conditional intervention improved the success rate under intermittent attacks, whereas continuously applying the same intervention substantially degraded policy performance.
arXiv cs.RO ·
Related topics
Interest trend
- Current
- 3
- Peak
- 4Nov 4 01:00
- Past 24 hours
- –
The trend compares participants observed throughout the same period, so its range may be smaller than the current interest count. Move or tap to inspect each hour; use the arrow keys on a keyboard.