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

2026-03-02

AI Summary

This work addresses the vulnerability of Video Large Language Models (VLLMs) to energy-latency denial-of-service attacks in safety-critical applications. Existing image-based adversarial methods fail due to temporal aggregation mechanisms and cannot accommodate the per-sample optimization constraints of real-time video streams. To overcome these limitations, we propose the first general-purpose denial-of-service attack framework tailored for VLLMs, introducing a novel adversarial trigger optimization paradigm that requires neither instance-specific tuning nor gradient computation during inference. Our approach leverages masked teacher forcing to induce high-cost outputs, augmented with rejection penalties and early-stopping suppression to mitigate temporal dilution and satisfy real-time constraints. Experiments demonstrate that our method triggers over 205× token inflation and more than 15× inference latency across three mainstream VLLMs, leading to severe safety violations in autonomous driving simulations.

Technology Category

• Computer Vision: Adversarial Attacks & Robustness
• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models

Application Category

• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

Video-LLMs are increasingly deployed in safety-critical applications but are vulnerable to Energy-Latency Attacks (ELAs) that exhaust computational resources. Current image-centric methods fail because temporal aggregation mechanisms dilute individual frame perturbations. Additionally, real-time demands make instance-wise optimization impractical for continuous video streams. We introduce VidDoS, which is the first universal ELA framework tailored for Video-LLMs. Our method leverages universal optimization to create instance-agnostic triggers that require no inference-time gradient calculation. We achieve this through $\textit{masked teacher forcing}$ to steer models toward expensive target sequences, combined with a $\textit{refusal penalty}$ and $\textit{early-termination suppression}$ to override conciseness priors. Testing across three mainstream Video-LLMs and three video datasets, which include video question answering and autonomous driving scenarios, shows extreme degradation. VidDoS induces a token expansion of more than 205$\times$ and inflates the inference latency by more than 15$\times$ relative to clean baselines. Simulations of real-time autonomous driving streams further reveal that this induced latency leads to critical safety violations. We urge the community to recognize and mitigate these high-hazard ELA in Video-LLMs.

Problem

• Video-LLMs
• Denial-of-Service
• Energy-Latency Attacks
• Temporal Aggregation
• Real-time Video Streams

Innovation

• universal denial-of-service attack
• video-based large language models
• energy-latency attack
• masked teacher forcing
• early-termination suppression