HalluProbe: Exploiting VLM Hallucinations for Backdoor Detection
Authors: Ji Guo, Xiaolong Qin, Jin Jiang, Yijing Liu, Pei Ke, Aiguo Chen, Wenbo Jiang
Organizations: Laboratory of Intelligent Collaborative Computing, University of Electronic Science and Technology of China · School of Software Engineering, Chengdu University of Information Technology · National Key Laboratory of Wireless Communications, University of Electronic Science and Technology of China · School of Computer Science and Engineering, University of Electronic Science and Technology of China
Vision-Language Models (VLMs) are vulnerable to backdoor attacks introduced through data poisoning, making effective backdoor detection a critical security challenge. In this paper, we uncover a novel connection between VLM hallucinations and backdoor behaviors and exploit this connection for backdoor detection. Our key insight is that poisoned training data induces strong language priors toward the backdoor target token, while such priors are also a major source of hallucinations. This suggests that enhancing hallucinations can expose hidden backdoor-related priors. Specifically, backdoored VLMs tend to hallucinate backdoor target tokens more frequently across diverse inputs, while the hallucinated tokens of clean VLMs are largely random and exhibit no stable target-specific bias. Based on this observation, we propose HalluProbe, which enhances hallucinations during output decoding and measures the distributional consistency of hallucinated tokens to identify backdoored VLMs. We evaluate HalluProbe against six representative VLM backdoor attacks across three models. Experimental results demonstrate that HalluProbe can effectively detect backdoored VLMs with an accuracy exceeding 90%, outperforming existing backdoor detection methods.
Figures & tables
Figure 1: An illustration of EHCD. The hallucinated token “dog” is intentionally amplified during generation by contrasting the output distribution with one that suppresses hallucinations.
Method
Training Data
Extra Training
Triggered Input
BYE [ 37 ]
✓
✗
✓
TCAP [ 31 ]
✓
✗
✓
SRD [ 42 ]
✓
✓
✓
ToT [ 36 ]
✗
✗
✓
HalluProbe (Ours)
✗
✗
✗
Table 1: Comparison of representative VLM backdoor defenses. “Training Data” indicates whether the method requires access to the fine-tuning dataset, and “Triggered Input” denotes whether a trigger-activated input is required during detection.
Figure 2: Illustration of the strengthened language prior induced by backdoor training. The backdoored VLM assigns an abnormally high logit to the attack target token “dog,” promoting it to a top candidate despite its semantic mismatch with the image.
Figure 3: Overview of HalluProbe. HalluProbe generates hallucination-enhanced and hallucination-suppressed outputs using EHCD and VCD, respectively. It then extracts and aggregates discrepant tokens across multiple images to construct a candidate token distribution. Finally, it calculates the frequency gap between the top two tokens and compares it against a predefined threshold to determine whether the examined VLM is backdoored.
Figure 4: Qualitative outputs of a backdoored VLM on probe images under VCD and EHCD. With hallucinations suppressed, the model describes the images faithfully; once amplified, the target token “dog” recurs across semantically unrelated images.
Figure 5: Distribution of HalluProbe scores S(M) over clean and backdoored checkpoints for the three victim models.
Figure 6: Detection accuracy of HalluProbe as a function of the threshold τ across six backdoor attacks and three victim backbones.
Figure 7: Detection performance of HalluProbe on Qwen3-VL-8B under BadNet with respect to the contrastive strength λ , the number of probe images N , and the noise step t .
Figure 8: The same probe image under increasing Gaussian noise.
T
Acc
P
R
F1
16
75.00
100.00
50.00
66.67
32
89.58
100.00
79.17
88.37
64
95.83
92.31
100.00
96.00
128
95.83
92.31
100.00
96.00
256
95.83
92.31
100.00
96.00
Table 4: Detection performance of HalluProbe as a function of the maximum decoding length T .
Figure 9: ASR and detection accuracy as functions of the BadNet poisoning rate across three victim models.
Figure 10: ASR and detection accuracy as functions of attack-training epochs under BadNet across three victim models.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 11: ASR and the detection accuracy of HalluProbe on the poisoned checkpoints as functions of the poisoning rate on Qwen3-VL across six attacks.
Figure 12: ASR and the detection accuracy of HalluProbe on the poisoned checkpoints as functions of the number of attack-training epochs on Qwen3-VL across six attacks.
Figure 13: Coverage count of the top-ten next-token candidates under EHCD, VCD, and their difference, and the ASR on each planted target over attack-training epochs under multiple-target poisoning.
Figure 14: Logit score rankings of a backdoored model under greedy decoding, EHCD, and VCD on four example poisoned images.
Figure 15: Detection accuracy of HalluProbe on LLaVA-NeXT as a function of the number of probe images N under each of the six attacks.
Figure 16: Coverage count of the top-ten next-token candidates under greedy decoding, EHCD, and VCD on three poisoned samples, together with the difference between EHCD and VCD.
Recent advances in Vision-Language Models (VLMs) have greatly enhanced the integration of visual perception and linguistic reasoning, driving rapid progress in multimodal understanding. Despite these achievements, the security of VLMs, particularly their vulnerability to backdoor attacks, remains significantly underexplored. Existing backdoor attacks on VLMs are still in an early stage of development, with most current methods relying on generating poisoned responses that contain fixed, easily identifiable patterns. In this work, we make two key contributions. First, we demonstrate for the first time that the stealthiness of existing VLM backdoor attacks has been substantially overestimated. By adapting defense techniques originally designed for other domains (e.g., vision-only and text-only models), we show that several state-of-the-art attacks can be detected with surprising ease. Second, to address this gap, we introduce Phantasia, a context-adaptive backdoor attack that dynamically aligns its poisoned outputs with the semantics of each input. Instead of producing static poisoned patterns, Phantasia encourages models to generate contextually coherent yet malicious responses that remain plausible, thereby significantly improving stealth and adaptability. Extensive experiments across diverse VLM architectures reveal that Phantasia achieves state-of-the-art attack success rates while maintaining benign performance under various defensive settings.
Nam Duong Tran, Phi Le Nguyen
Institute for AI Innovation and Societal Impact, Hanoi University of Science and Technology
Large Vision-Language Models (LVLMs) have demonstrated remarkable capabilities across various tasks, yet they remain vulnerable to backdoor attacks. Existing defense methods predominantly focus on sample-level defense, which relies on the knowledge of training data or triggers. However, identifying whether a given model is backdoored remains a critical but unexplored task. To fill this gap, we propose EntropyScan, a lightweight and trigger-agnostic method for model-level backdoor detection in LVLMs. We first observe that backdoor injection disrupts the cross-modal alignment, resulting in pronounced structural anomalies in visual attention allocation on benign samples. Based on this insight, EntropyScan detects the backdoor models by quantifying such attention deviations. Specifically, it extracts visual attention distributions from the initial layers of the Large Language Model (LLM) and applies Tsallis entropy to capture these structural distortions. By employing a reference-anchored Z-score normalization on a small set of benign samples, it effectively identifies the backdoored model. Extensive experiments across two LVLMs architectures and three advanced attack scenarios show that EntropyScan achieves an F1 score of 98.5% in average and an AUC of 96.6%. Our code will be publicly available soon.
Xuanyu Ge, Zhongqi Wang, Jie Zhang +2
China University of Geosciences · University of the Chinese Academy of Sciences · Institute of Computing Technology, Chinese Academy of Sciences
Backdoor attacks undermine the reliability and trustworthiness of machine learning systems by injecting hidden behaviors that can be maliciously activated at inference time. While such threats have been extensively studied in unimodal settings, their impact on multimodal foundation models, particularly vision-language models (VLMs), remains largely underexplored. In this work, we introduce BackdoorVLM, the first comprehensive benchmark for evaluating backdoor attacks and defenses on VLMs across a broad range of settings. It adopts a unified perspective that injects and analyzes backdoors across core vision-language tasks, including image captioning and visual question answering. BackdoorVLM organizes multimodal backdoor threats into 5 representative categories: targeted refusal, malicious injection, jailbreak, concept substitution, and perceptual hijack. Each category captures a distinct pathway through which an adversary can manipulate a model's behavior. We evaluate these threats using 12 representative attack methods spanning text, image, and bimodal triggers, tested on 2 open-source VLMs and 3 multimodal datasets along with 5 representative defense methods. Our analysis reveals that VLMs exhibit strong sensitivity to textual instructions, and in bimodal backdoors the text trigger typically overwhelms the image trigger when forming the backdoor mapping. Notably, backdoors involving the textual modality remain highly potent, with poisoning rates as low as 1% yielding over 90% success across most tasks. Moreover, existing defenses lack generalization across trigger types and backdoor scenarios, leading to unreliable protection against diverse triggers across modalities and forms. These findings highlight significant, underexplored vulnerabilities in current VLMs. We hope that BackdoorVLM can serve as a useful benchmark for analyzing and mitigating multimodal backdoor threats.
Juncheng Li, Yige Li, Hanxun Huang +5
Fudan University · Singapore Management University · The University of Melbourne