VidDoS: Universal Denial-of-Service Attack on Video-based Large Language Models

Researchers have discovered VidDoS, a new universal attack framework that can severely degrade Video-based Large Language Models by causing extreme computational resource exhaustion. The attack increases token generation by over 205x and inference latency by more than 15x, creating critical safety risks in real-world applications like autonomous driving.

• →VidDoS is the first universal Energy-Latency Attack framework specifically designed to target Video-LLMs.
• →The attack causes extreme performance degradation with token expansion over 205x and latency increases over 15x compared to normal operations.
• →Testing across three mainstream Video-LLMs and datasets showed consistent vulnerability across video question answering and autonomous driving scenarios.
• →Real-time autonomous driving simulations revealed that the induced latency leads to critical safety violations.
• →The attack uses masked teacher forcing and optimization techniques that require no inference-time gradient calculation, making it practically deployable.