Research map / Adversarial robustness and security / Backdoor and poisoning
BYORn: Bootstrap Your Own Responses to Defend Large Vision-Language Models Against Backdoor Attacks
Ivan Sabolic, Marin Oršić, Josip Šarić, Sven Loncaric
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 |
|---|---|---|---|
| regenerate suspicious responses with EMA model instead of filtering them (BYORn vs BYORn-F) (ablation) | BYORn · Blend, p=0.975, LLaVA | ASR (%): 79.3 -> 1.2 | Improved |
| regenerate suspicious responses instead of filtering (ablation) | BYORn · Blend, p=0.95, LLaVA | ASR (%): 57.6 -> 1.1 | Improved |
| regenerate responses instead of filtering (ablation) | BYORn · Image captioning Flickr, LLaVA | CIDEr: 42.4 -> 62.0 | Improved |
| regenerate responses instead of filtering (ablation) | BYORn · Spot the Difference, LLaVA | CIDEr: 129.8 -> 134.9 | Improved |
| add poison-aware minibatching (ablation) | BYORn · LLaVA training on LADD | epoch time (min): 301 -> 135 | Improved |
| add speculative decoding with a LLaVA-OneVision draft model | BYORn · LLaVA training on LADD | epoch time (min): 301 -> 294 | No clear effect |
| raise percentile threshold p from 0.6 to 0.975 (ablation) | BYORn · Blend, LLaVA | ASR (%): 0.0 -> 1.2 | No clear effect |
| lower p to 0.3 (many false positives) (ablation) | BYORn · Blend, LLaVA | CIDEr: 62.2 -> 55.6 | Hurt |
| vary poisoning rate from 1% to 50% (ablation) | BYORn · VL-Trojan, LLaVA Flickr30k | ASR (%): 0.0 -> 0.2 | No clear effect |
| vary EMA decay lambda from 0.99 to 0.999 (ablation) | BYORn · Blend, LLaVA | CIDEr: 61.8 -> 61.2 | No clear effect |
| apply BYORn defense against adaptive image-wide attack | LLaVA SFT · Adaptive image-wide, COCO | ASR (%): 79.4 -> 8.4 | Improved |
| apply BYORn against PGD adversarial-perturbation trigger | LLaVA SFT · Flickr30k | ASR (%): 92.0 -> 0.2 | Improved |
Closest papers by meaning
- Revisiting Backdoor Attacks against Large Vision-Language Models from Domain Shift (CVPR 2025 · similarity 0.91)
- Semantic Shield: Defending Vision-Language Models Against Backdooring and Poisoning via Fine-grained Knowledge Alignment (CVPR 2024 · similarity 0.89)
- CROW: Eliminating Backdoors from Large Language Models via Internal Consistency Regularization (ICML 2025 · similarity 0.88)
- BadVLA: Towards Backdoor Attacks on Vision-Language-Action Models via Objective-Decoupled Optimization (NeurIPS 2025 · similarity 0.88)
- Steering Away from Harm: An Adaptive Approach to Defending Vision Language Model Against Jailbreaks (CVPR 2025 · similarity 0.88)
- 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.
