safety security
The Robot Defense Worked Only When It Knew When to Intervene
A sparse-autoencoder feature triggered selective suppression of adversarial patches in a vision-language-action policy.
Summary
A sparse-autoencoder feature triggered selective suppression of adversarial patches in a vision-language-action policy.
Researchers identified an internal VLA feature correlated with adversarial patches, then suppressed it at inference time only when a linear probe detected an attack. On LIBERO-10, conditional intervention improved success under intermittent attacks, while continuously suppressing the same feature substantially harmed ordinary policy performance. The study tests one mechanistic defense setting and does not establish broad patch immunity.
Why it matters
A sparse-autoencoder feature triggered selective suppression of adversarial patches in a vision-language-action policy.
Limits and context
- Researchers identified an internal VLA feature correlated with adversarial patches, then suppressed it at inference time only when a linear probe detected an attack.
- The study tests one mechanistic defense setting and does not establish broad patch immunity.
Key claims
A sparse-autoencoder feature triggered selective suppression of adversarial patches in a vision-language-action policy.
Qualification: Researchers identified an internal VLA feature correlated with adversarial patches, then suppressed it at inference time only when a linear probe detected an attack.
Evidence: source-2026-10-05-008
Sources
- arXiv preprint 2610.03498arXiv · primary research
Corrections
No corrections have been recorded for this story.