Research map / Adversarial robustness and security / Backdoor and poisoning
Probing Semantic Insensitivity for Inference-Time Backdoor Defense in Multimodal Large Language Model
Xuankun Rong, Wenke Huang, Wenzheng Jiang, Yiming Li, Wenxuan Wang, Mang Ye
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 ToT inference-time detection (vs ZIP, best baseline) to backdoored LLaVA | LLaVA-v1.5-7B · ScienceQA (BadNet) | ASR (%): 98.66 -> 0.00 | Improved |
| add ToT defense to backdoored InternVL | InternVL2.5-8B · IconQA (BadNet) | ASR (%): 94.49 -> 0.30 | Improved |
| add ToT defense; clean accuracy preserved | LLaVA-v1.5-7B · ScienceQA (BadNet) | CP (%): 90.23 -> 90.22 | No clear effect |
| apply image-purification defenses (DiffPure, ZIP, SampDetox) instead of ToT | InternVL2.5-8B · ScienceQA (BadNet) | CP (%): 98.07 -> 86.07 | Hurt |
| add Textual Semantic Consistency (TSC) module to CDG-only detector (ablation) | ToT · ScienceQA/IconQA, BadNet/Blended | TP | Improved |
| add Confidence Drift Gap (CDG) module to TSC-only detector (ablation) | ToT · ScienceQA/IconQA, BadNet/Blended | TP | Improved |
| add both TSC and CDG vs neither (ablation) | ToT · ScienceQA/IconQA, BadNet/Blended | TP | Improved |
| lower poison rate from 10% to 1% (BadNet) (ablation) | ToT · ScienceQA (BadNet) | ASR (%): 0.00 -> 4.77 | Mixed |
| lower poison rate from 10% to 1% (Blended) (ablation) | ToT · ScienceQA (Blended) | ASR (%): 0.00 -> 1.67 | Mixed |
| attacker uses paraphrased target responses (Variant Response) against ToT | ToT · ScienceQA (BadNet, Variant Response) | ASR (%): 97.67 -> 3.78 | Improved |
| attacker adds text trigger word (Multimodal Trigger) against ToT | ToT · ScienceQA (Blended, Multimodal Trigger) | ASR (%): 21.76 -> 0.69 | Improved |
Closest papers by meaning
- BackdoorLLM: A Comprehensive Benchmark for Backdoor Attacks and Defenses on Large Language Models (NeurIPS 2025 · similarity 0.91)
- Revisiting Backdoor Attacks against Large Vision-Language Models from Domain Shift (CVPR 2025 · similarity 0.91)
- Backdoor Cleaning without External Guidance in MLLM Fine-tuning (NeurIPS 2025 · similarity 0.90)
- BadToken: Token-level Backdoor Attacks to Multi-modal Large Language Models (CVPR 2025 · similarity 0.89)
- CROW: Eliminating Backdoors from Large Language Models via Internal Consistency Regularization (ICML 2025 · similarity 0.89)
- Backdooring Vision-Language Models with Out-Of-Distribution Data (ICLR 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.
