Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·Aug 24·5 min read
CIVA: Critic-Induced Value-Subspace Attacks on Visual World-Model Agents
TL;DR
CIVA introduces a novel attack method targeting visual world-model agents like DreamerV3 by exploiting low-dimensional value subspaces induced by the agent's critic. This approach significantly outperforms existing methods, achieving a 26.07% reward drop in specific environments while maintaining temporal coherence.
✦ Why It Matters
Consider implementing CIVA techniques to enhance your adversarial testing against visual world-model agents.
Key Takeaways
- Explore CIVA for attacking visual world-model agents.
- Utilize SVD to identify low-dimensional value subspaces.
- Implement exponential moving averages for smoother perturbations.
- Test CIVA against various environments for robust evaluation.
How It Works
CIVA operates by first probing the target agent offline to identify a low-dimensional subspace of value perturbations using Projected Gradient Descent (PGD) and Singular Value Decomposition (SVD). During the attack, it optimizes only the coefficients of this subspace, which allows for efficient and coherent perturbations that exploit the agent's recurrent dynamics.