Detect and Suppress runs a sparse autoencoder on a VLA’s internal representations and finds a feature whose activation tracks the presence of an adversarial patch. A linear probe flags an attack, and only then is that feature suppressed at inference time, with no fine-tuning. On LIBERO-10 the conditional intervention improves success under intermittent attacks, while applying the same suppression all the time hurts the policy badly.
Relevance
The useful result is about timing: an always-on defence costs nominal performance, so a detector has to decide when to act. That makes it a runtime monitoring problem as much as a robot security one, for robust robotics.
Links
- arXiv
- Subfield: Security, Robot learning, Perception