Dual-Modality Multi-Stage Adversarial Safety Training: Robustifying Multimodal Web Agents Against Cross-Modal Attacks

• URL: http://arxiv.org/abs/2603.04364v1
• Date: Wed, 04 Mar 2026 18:29:54 GMT
• Title: Dual-Modality Multi-Stage Adversarial Safety Training: Robustifying Multimodal Web Agents Against Cross-Modal Attacks
• Authors: Haoyu Liu, Dingcheng Li, Lukas Rutishauser, Zeyu Zheng,
• Abstract summary: We propose a framework that formalizes the agent-attacker interaction as a two-player zero-sum Markov game and co-trains both players through a three-stage pipeline.<n>Our approach significantly outperforms established training-based and prompt-based defenses.
• Score: 23.881766496924502
• License: http://creativecommons.org/licenses/by-nc-sa/4.0/
• Abstract: Multimodal web agents that process both screenshots and accessibility trees are increasingly deployed to interact with web interfaces, yet their dual-stream architecture opens an underexplored attack surface: an adversary who injects content into the webpage DOM simultaneously corrupts both observation channels with a consistent deceptive narrative. Our vulnerability analysis on MiniWob++ reveals that attacks including a visual component far outperform text-only injections, exposing critical gaps in text-centric VLM safety training. Motivated by this finding, we propose Dual-Modality Multi-Stage Adversarial Safety Training (DMAST), a framework that formalizes the agent-attacker interaction as a two-player zero-sum Markov game and co-trains both players through a three-stage pipeline: (1) imitation learning from a strong teacher model, (2) oracle-guided supervised fine-tuning that uses a novel zero-acknowledgment strategy to instill task-focused reasoning under adversarial noise, and (3) adversarial reinforcement learning via Group Relative Policy Optimization (GRPO) self-play. On out-of-distribution tasks, DMAST substantially mitigates adversarial risks while simultaneously doubling task completion efficiency. Our approach significantly outperforms established training-based and prompt-based defenses, demonstrating genuine co-evolutionary progress and robust generalization to complex, unseen environments.

Related papers

• Adversarial Attack-Defense Co-Evolution for LLM Safety Alignment via Tree-Group Dual-Aware Search and Optimization [51.12422886183246]
  Large Language Models (LLMs) have developed rapidly in web services, delivering unprecedented capabilities while amplifying societal risks.<n>Existing works tend to focus on either isolated jailbreak attacks or static defenses, neglecting the dynamic interplay between evolving threats and safeguards in real-world web contexts.<n>We propose ACE-Safety, a novel framework that jointly optimize attack and defense models by seamlessly integrating two key innovative procedures.
  arXiv  Detail & Related papers  (2025-11-24T15:23:41Z)
• Debiased Dual-Invariant Defense for Adversarially Robust Person Re-Identification [52.63017280231648]
  Person re-identification (ReID) is a fundamental task in many real-world applications such as pedestrian trajectory tracking.<n>Person ReID models are highly susceptible to adversarial attacks, where imperceptible perturbations to pedestrian images can cause entirely incorrect predictions.<n>We propose a dual-invariant defense framework composed of two main phases.
  arXiv  Detail & Related papers  (2025-11-13T03:56:40Z)
• Visual Backdoor Attacks on MLLM Embodied Decision Making via Contrastive Trigger Learning [89.1856483797116]
  We introduce BEAT, the first framework to inject visual backdoors into MLLM-based embodied agents.<n>Unlike textual triggers, object triggers exhibit wide variation across viewpoints and lighting, making them difficult to implant reliably.<n>BEAT achieves attack success rates up to 80%, while maintaining strong benign task performance.
  arXiv  Detail & Related papers  (2025-10-31T16:50:49Z)
• Neutral Agent-based Adversarial Policy Learning against Deep Reinforcement Learning in Multi-party Open Systems [3.431456142488844]
  We propose a neutral agent-based approach across various task scenarios in multi-party open systems.<n>We evaluate our proposed method on the SMAC platform based on Starcraft II and the autonomous driving simulation platform Highway-env.
  arXiv  Detail & Related papers  (2025-10-13T02:53:22Z)
• Adversarial Reinforcement Learning for Large Language Model Agent Safety [20.704989548285372]
  Large Language Model (LLM) agents can leverage tools like Google Search to complete complex tasks.<n>Current defense strategies rely on fine-tuning LLM agents on datasets of known attacks.<n>We propose Adversarial Reinforcement Learning for Agent Safety (ARLAS), a novel framework that leverages adversarial reinforcement learning (RL) by formulating the problem as a two-player zero-sum game.
  arXiv  Detail & Related papers  (2025-10-06T23:09:18Z)
• AdvEvo-MARL: Shaping Internalized Safety through Adversarial Co-Evolution in Multi-Agent Reinforcement Learning [78.5751183537704]
  AdvEvo-MARL is a co-evolutionary multi-agent reinforcement learning framework that internalizes safety into task agents.<n>Rather than relying on external guards, AdvEvo-MARL jointly optimize attackers and defenders.
  arXiv  Detail & Related papers  (2025-10-02T02:06:30Z)
• DINA: A Dual Defense Framework Against Internal Noise and External Attacks in Natural Language Processing [12.279803315688218]
  Large language models (LLMs) and generative AI become increasingly integrated into customer service and moderation applications.<n>In this work, we identify and systematically address these dual adversarial threats by introducing DINA (Dual Defense Against Internal Noise and Adversarial Attacks)<n>Our approach adapts advanced noisy-label learning methods from computer vision and integrates them with adversarial training to simultaneously mitigate internal label sabotage and external adversarial perturbations.
  arXiv  Detail & Related papers  (2025-08-04T16:33:17Z)
• Manipulating Multimodal Agents via Cross-Modal Prompt Injection [34.35145839873915]
  We identify a critical yet previously overlooked security vulnerability in multimodal agents.<n>We propose CrossInject, a novel attack framework in which attackers embed adversarial perturbations across multiple modalities.<n>Our method outperforms state-of-the-art attacks, achieving at least a +30.1% increase in attack success rates.
  arXiv  Detail & Related papers  (2025-04-19T16:28:03Z)
• Few-Shot Adversarial Prompt Learning on Vision-Language Models [62.50622628004134]
  The vulnerability of deep neural networks to imperceptible adversarial perturbations has attracted widespread attention. Previous efforts achieved zero-shot adversarial robustness by aligning adversarial visual features with text supervision. We propose a few-shot adversarial prompt framework where adapting input sequences with limited data makes significant adversarial robustness improvement.
  arXiv  Detail & Related papers  (2024-03-21T18:28:43Z)
• Mutual-modality Adversarial Attack with Semantic Perturbation [81.66172089175346]
  We propose a novel approach that generates adversarial attacks in a mutual-modality optimization scheme. Our approach outperforms state-of-the-art attack methods and can be readily deployed as a plug-and-play solution.
  arXiv  Detail & Related papers  (2023-12-20T05:06:01Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.