Research map / Adversarial robustness and security / Backdoor and poisoning
Backdoor Attacks on Open Vocabulary Object Detectors via Multi-Modal Prompt Tuning
Ankita Raj, Chetan Arora
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 |
|---|---|---|---|
| tune prompts in both vision and text branches instead of text-only (CoCoOp-Det) | TrAP · ODinW Vehicles (OMA) | ASR: 0.08 -> 0.79 | Improved |
| add text prompts to image-only VPT | TrAP · ODinW Vehicles (OMA) | ASR: 0.64 -> 0.79 | Improved |
| add curriculum trigger-size shrinking vs training only at rho=0.1 (ablation) | TrAP · ODinW Vehicles (OMA) | ASR: 0.75 -> 0.79 | Improved |
| add curriculum vs train large (0.2) test small (0.1) (ablation) | TrAP · ODinW Vehicles (OMA) | ASR: 0.66 -> 0.79 | Improved |
| add image-conditioned meta-net to text prompt (ablation) | TrAP · ODinW Vehicles (OMA) | BmAP: 63.9 -> 64.9 | Improved |
| fine-tune model weights instead of prompt tuning (ablation) | TrAP · ODinW Vehicles (OMA) | BmAP / ASR: BmAP 64.9 -> 67.4; ASR 0.79 -> 0.74 (Fine-tune-A) | Mixed |
| increase trigger size to rho=0.5 (ablation) | TrAP · ODinW Vehicles (OMA) | ASR: 0.79 -> 0.90 | Improved |
| decrease trigger size to rho=0.05 (ablation) | TrAP · ODinW Vehicles (OMA) | ASR: 0.79 -> 0.75 | Hurt |
| apply PatchDrop defense (50%) | TrAP · ODinW Vehicles (OMA) | ASR: 0.79 -> 0.63; BmAP 64.9 -> 43.9 | Hurt |
| rephrase target class in prompt (Bus -> A Bus) | TrAP · ODinW Vehicles (OMA) | ASR: 0.79 -> 0.04 | Improved |
| apply PAD adversarial patch defense | TrAP (checkerboard trigger) · ODinW Vehicles (OMA) | ASR: 0.48 -> 0.50 | No clear effect |
| apply TrAP prompt tuning on GLIP-T instead of Grounding DINO | TrAP · ODinW (6 datasets, OMA) | ASR: 0.84 to 1.00 across datasets | Improved |
Closest papers by meaning
- Revisiting Backdoor Attacks against Large Vision-Language Models from Domain Shift (CVPR 2025 · similarity 0.89)
- SEER: Backdoor Detection for Vision-Language Models through Searching Target Text and Image Trigger Jointly (AAAI 2024 · similarity 0.87)
- Backdooring Vision-Language Models with Out-Of-Distribution Data (ICLR 2025 · similarity 0.87)
- BadCLIP: Trigger-Aware Prompt Learning for Backdoor Attacks on CLIP (CVPR 2024 · similarity 0.86)
- Defending Multimodal Backdoored Models by Repulsive Visual Prompt Tuning (NeurIPS 2025 · similarity 0.86)
- BadVLA: Towards Backdoor Attacks on Vision-Language-Action Models via Objective-Decoupled Optimization (NeurIPS 2025 · similarity 0.86)
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.
