Research map / Adversarial robustness and security / Attacks on text and graphs
VLMSysTrojan: Stealthy System-Aware Backdoor Attacks Against Vision-Language Models
Yezheng Cheng, Zexin Li, Jiaqi Wu, Zhihong Zhang, Simin Chen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add trigger optimization, guard-bias and decoder fine-tuning together (all three modules) (ablation) | VLMSysTrojan · Module ablation | ASR (%): 51.3 -> 99.8 | Improved |
| add trigger optimization and guard-bias to decoder fine-tuning (ablation) | VLMSysTrojan · Module ablation | SRR (%): 46.6 -> 100.0 | Improved |
| add decoder fine-tuning to trigger optimization only (ablation) | VLMSysTrojan · Module ablation | ASR (%): 0.0 -> 99.8 | Improved |
| add fine-tuning after guard-bias injection (ablation) | VLMSysTrojan · Module ablation | Pre-Func: 25.6 -> 49.1 | Improved |
| combine all four loss objectives l1-l4 (ablation) | VLMSysTrojan · Objective ablation | SRR and ASR: 100.0 and 99.8 | Improved |
| use system-level kernel-aware backdoor instead of model-level baselines | VLMSysTrojan vs VL-Trojan/Shadowcast · Pre-deployment stealthiness | SRR: 99.8% vs below 5% | Improved |
| deploy on different backend (CUDAGraphs, ONNXRT, TVM) | VLMSysTrojan · Flickr8k BLIP | ASR (%): 98.9, 99.3, 99.5 | No clear effect |
| apply quantization, Gaussian noise or JPEG defenses | VLMSysTrojan · Defense evaluation | ASR (%): 99.8 -> 99.2 | No clear effect |
| apply Neural Cleanse, SCAn, MM-BD detection | VLMSysTrojan · Backdoor detection | anomaly index: approximately 1.9 | No clear effect |
| change TorchInductor version V2.8, V2.9, V2.10 | VLMSysTrojan · Kernel versions | ASR (%): 99.8, 99.8, 99.8 | No clear effect |
| apply to QWen3.5-0.8B and QWen3.5-2B | VLMSysTrojan · Flickr8k | ASR (%): 99.6, 99.5 | No clear effect |
| append trigger to inputs | VLMSysTrojan · Functionality BLIP | CIDEr: 84.5 -> 81.1 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- VL-Trojan — VL-Trojan: Multimodal Instruction Backdoor Attacks against Autoregressive Visual Language Models (IJCV 2025) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Revisiting Backdoor Attacks against Large Vision-Language Models from Domain Shift (CVPR 2025 · similarity 0.90)
- MMJ-Bench: A Comprehensive Study on Jailbreak Attacks and Defenses for Vision Language Models (AAAI 2025 · similarity 0.90)
- Backdooring Vision-Language Models with Out-Of-Distribution Data (ICLR 2025 · similarity 0.89)
- BadVLA: Towards Backdoor Attacks on Vision-Language-Action Models via Objective-Decoupled Optimization (NeurIPS 2025 · similarity 0.89)
- Stealthy Backdoor Attack in Self-Supervised Learning Vision Encoders for Large Vision Language Models (CVPR 2025 · similarity 0.89)
- The VLLM Safety Paradox: Dual Ease in Jailbreak Attack and Defense (NeurIPS 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
