New CIVA attack method targets visual world-model agents

Researchers have developed a new method called Critic-Induced Value-Subspace Attacks (CIVA) to target visual world-model agents. These agents, like DreamerV3, operate using a recurrent latent state, making them resilient to traditional frame-by-frame attacks. CIVA leverages the agent's own critic to identify a low-dimensional subspace where perturbations are most effective. By optimizing within this subspace and smoothing the perturbations, CIVA achieves significant reward drops in environments such as DMC walker walk, Pong, and Crafter, outperforming existing attack methods. AI

IMPACT This research highlights potential vulnerabilities in visual world-model agents, suggesting a need for improved adversarial robustness in AI systems.

RANK_REASON The cluster contains a research paper detailing a novel attack method on AI agents. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Crafter
• Critic-Induced Value-Subspace Attacks
• DreamerV3
• Pong
• Projected Gradient Descent
• singular value decomposition

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →