Research map / Adversarial robustness and security / Backdoor and poisoning
TokenSwap: Backdoor Attack on the Compositional Understanding of Large Vision-Language Models
Zhifang Zhang, Qiqi Tao, Jiaqi Lv, Na Zhao, Lei Feng, Joey Tianyi Zhou
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 token swapping with adaptive token-weighted loss (TokenSwap vs BadNet) | BLIP2 · Flickr30k (poisoned input) | ASR (%): 53.13 -> 80.47 | Improved |
| add TokenSwap training instead of BadNet | Qwen-VL2.5-7B · Flickr30k (poisoned input) | ASR (%): 45.35 -> 73.04 | Improved |
| add TokenSwap training instead of BadNet | LLaVA-7B · Flickr30k (poisoned input) | ASR (%): 78.91 -> 85.16 | Improved |
| add TokenSwap training instead of BadNet | BLIP2 trained on MSCOCO · MSCOCO->Flickr8k (cross-dataset) | ASR (%): 54.69 -> 88.28 | Improved |
| add TokenSwap loss instead of TrojVLM/MABA fixed-target attacks | BLIP2 · Flickr8k (poisoned input, attack comparison) | ASR (%): 80.47 -> 91.41 | Improved |
| add adaptive token-weighted emphasis (alpha > 0) (ablation) | TokenSwap · Flickr8k ablation (alpha and gamma) | ASR (%) | Improved |
| add focal-style gamma weighting (ablation) | TokenSwap · Flickr8k ablation (alpha and gamma) | ASR (%) | Improved |
| raise poisoning rate from 0.1 to 0.5 (ablation) | TokenSwap · Flickr8k ablation (poisoning rate) | ASR (%): 75.00 -> 91.41 | Improved |
| raise poisoning rate from 0.5 to 0.9 (ablation) | TokenSwap · Flickr8k ablation (poisoning rate) | ASR (%): 91.41 -> 92.97 | No clear effect |
| vary trigger type (Black to Blended noise) (ablation) | TokenSwap · Flickr8k ablation (trigger type) | ASR (%): 87.50 -> 92.19 | No clear effect |
| vary trigger size from 20 to 50 (ablation) | TokenSwap · Flickr8k ablation (trigger size) | ASR (%): 89.84 -> 90.63 | No clear effect |
| apply TokenSwap to VQA instead of BadNet | BLIP2-style VQA models · VQAv2 (poisoned input) | ASR (%): 79.89 -> 84.83 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- VL-Trojan — VL-Trojan: Multimodal Instruction Backdoor Attacks against Autoregressive Visual Language Models (IJCV 2025) · better on 1 of 1 shared result
- TrojVLM — TrojVLM: Backdoor Attack Against Vision Language Models (ECCV 2024) · better on 1 of 1 shared result
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 0.93)
- Backdooring Vision-Language Models with Out-Of-Distribution Data (ICLR 2025 · similarity 0.91)
- Semantic Shield: Defending Vision-Language Models Against Backdooring and Poisoning via Fine-grained Knowledge Alignment (CVPR 2024 · similarity 0.91)
- FigStep: Jailbreaking Large Vision-Language Models via Typographic Visual Prompts (AAAI 2025 · similarity 0.90)
- SLADE: Shielding against Dual Exploits in Large Vision-Language Models (CVPR 2025 · similarity 0.89)
- VERA-V: Variational Inference Framework for Jailbreaking Vision-Language Models (ICML 2026 · similarity 0.89)
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.
