Test-Time Attention Purification for Backdoored Large Vision Language Models

• URL: http://arxiv.org/abs/2603.12989v1
• Date: Fri, 13 Mar 2026 13:45:06 GMT
• Title: Test-Time Attention Purification for Backdoored Large Vision Language Models
• Abstract summary: We provide a new mechanistic understanding of backdoor behaviors in large vision-language models (LVLMs)<n>We propose CleanSight, a training-free, plug-and-play defense that operates purely at test time.<n>CleanSight significantly outperforms existing pixel-based purification defenses across diverse datasets and backdoor attack types.
• Score: 23.890959327899925
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Despite the strong multimodal performance, large vision-language models (LVLMs) are vulnerable during fine-tuning to backdoor attacks, where adversaries insert trigger-embedded samples into the training data to implant behaviors that can be maliciously activated at test time. Existing defenses typically rely on retraining backdoored parameters (e.g., adapters or LoRA modules) with clean data, which is computationally expensive and often degrades model performance. In this work, we provide a new mechanistic understanding of backdoor behaviors in LVLMs: the trigger does not influence prediction through low-level visual patterns, but through abnormal cross-modal attention redistribution, where trigger-bearing visual tokens steal attention away from the textual context - a phenomenon we term attention stealing. Motivated by this, we propose CleanSight, a training-free, plug-and-play defense that operates purely at test time. CleanSight (i) detects poisoned inputs based on the relative visual-text attention ratio in selected cross-modal fusion layers, and (ii) purifies the input by selectively pruning the suspicious high-attention visual tokens to neutralize the backdoor activation. Extensive experiments show that CleanSight significantly outperforms existing pixel-based purification defenses across diverse datasets and backdoor attack types, while preserving the model's utility on both clean and poisoned samples.

Related papers

• Self-Purification Mitigates Backdoors in Multimodal Diffusion Language Models [74.1970982768771]
  We show that well-established data-poisoning pipelines can successfully implant backdoors into MDLMs.<n>We introduce a backdoor defense framework for MDLMs named DiSP (Diffusion Self-Purification)
  arXiv  Detail & Related papers  (2026-02-24T15:47:52Z)
• Backdoor Attacks on Open Vocabulary Object Detectors via Multi-Modal Prompt Tuning [5.0734761482919115]
  Open-vocabulary object detectors (OVODs) unify vision and language to detect arbitrary object categories based on text prompts.<n>We conduct the first study of backdoor attacks on OVODs and reveal a new attack surface introduced by prompt tuning.
  arXiv  Detail & Related papers  (2025-11-16T19:05:31Z)
• Sealing The Backdoor: Unlearning Adversarial Text Triggers In Diffusion Models Using Knowledge Distillation [3.54387829918311]
  Adversaries can inject imperceptible textual triggers into training data, causing models to generate manipulated outputs.<n>We propose Self-Knowledge Distillation with Cross-Attention Guidance (SKD-CAG) to erase associations between adversarial text triggers and poisoned outputs.<n>Our method achieves removal accuracy 100% for pixel backdoors and 93% for style-based attacks, without sacrificing robustness or image fidelity.
  arXiv  Detail & Related papers  (2025-08-20T00:57:21Z)
• Poison Once, Control Anywhere: Clean-Text Visual Backdoors in VLM-based Mobile Agents [54.35629963816521]
  This work introduces VIBMA, the first clean-text backdoor attack targeting VLM-based mobile agents.<n>The attack injects malicious behaviors into the model by modifying only the visual input.<n>We show that our attack achieves high success rates while preserving clean-task behavior.
  arXiv  Detail & Related papers  (2025-06-16T08:09:32Z)
• Backdoor Cleaning without External Guidance in MLLM Fine-tuning [76.82121084745785]
  Believe Your Eyes (BYE) is a data filtering framework that leverages attention entropy patterns as self-supervised signals to identify and filter backdoor samples.<n>It achieves near-zero attack success rates while maintaining clean-task performance.
  arXiv  Detail & Related papers  (2025-05-22T17:11:58Z)
• Neural Antidote: Class-Wise Prompt Tuning for Purifying Backdoors in CLIP [51.04452017089568]
  Class-wise Backdoor Prompt Tuning (CBPT) is an efficient and effective defense mechanism that operates on text prompts to indirectly purify CLIP.<n>CBPT significantly mitigates backdoor threats while preserving model utility.
  arXiv  Detail & Related papers  (2025-02-26T16:25:15Z)
• Proactive Adversarial Defense: Harnessing Prompt Tuning in Vision-Language Models to Detect Unseen Backdoored Images [0.0]
  Backdoor attacks pose a critical threat by embedding hidden triggers into inputs, causing models to misclassify them into target labels.<n>We introduce a groundbreaking method to detect unseen backdoored images during both training and inference.<n>Our approach trains learnable text prompts to differentiate clean images from those with hidden backdoor triggers.
  arXiv  Detail & Related papers  (2024-12-11T19:54:14Z)
• Pre-trained Trojan Attacks for Visual Recognition [106.13792185398863]
  Pre-trained vision models (PVMs) have become a dominant component due to their exceptional performance when fine-tuned for downstream tasks. We propose the Pre-trained Trojan attack, which embeds backdoors into a PVM, enabling attacks across various downstream vision tasks. We highlight the challenges posed by cross-task activation and shortcut connections in successful backdoor attacks.
  arXiv  Detail & Related papers  (2023-12-23T05:51:40Z)
• BadCLIP: Dual-Embedding Guided Backdoor Attack on Multimodal Contrastive Learning [85.2564206440109]
  This paper reveals the threats in this practical scenario that backdoor attacks can remain effective even after defenses. We introduce the emphtoolns attack, which is resistant to backdoor detection and model fine-tuning defenses.
  arXiv  Detail & Related papers  (2023-11-20T02:21:49Z)
• Untargeted Backdoor Attack against Object Detection [69.63097724439886]
  We design a poison-only backdoor attack in an untargeted manner, based on task characteristics. We show that, once the backdoor is embedded into the target model by our attack, it can trick the model to lose detection of any object stamped with our trigger patterns.
  arXiv  Detail & Related papers  (2022-11-02T17:05:45Z)
• Training set cleansing of backdoor poisoning by self-supervised representation learning [0.0]
  A backdoor or Trojan attack is an important type of data poisoning attack against deep neural network (DNN) We show that supervised training may build stronger association between the backdoor pattern and the associated target class than that between normal features and the true class of origin. We propose to use unsupervised representation learning to avoid emphasising backdoor-poisoned training samples and learn a similar feature embedding for samples of the same class.
  arXiv  Detail & Related papers  (2022-10-19T03:29:58Z)
• Invisible Backdoor Attacks Using Data Poisoning in the Frequency Domain [8.64369418938889]
  We propose a generalized backdoor attack method based on the frequency domain. It can implement backdoor implantation without mislabeling and accessing the training process. We evaluate our approach in the no-label and clean-label cases on three datasets.
  arXiv  Detail & Related papers  (2022-07-09T07:05:53Z)

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.