Research map / Adversarial robustness and security / Attacks on text and graphs
When and Where do Data Poisons Attack Textual Inversion?
Jeremy Styborski, Mingzhi Lyu, Jiayou Lu, Nupur Kapur, Adams Wai-Kin Kong
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 JPEG compression (quality 25) as a defense against poisons | Textual Inversion on CustomConcept101 · CustomConcept101 poison average | DINOv2 similarity: 0.21 -> 0.37 | Improved |
| restrict TI training to high timesteps (T600) (ablation) | Textual Inversion on CustomConcept101 · CustomConcept101 poison average | DINOv2 similarity: 0.21 -> 0.22 (ADM+ 0.10 -> 0.25, ADM- 0.33 -> 0.27) | Mixed |
| add loss masking to constrain learning to concept regions (ablation) | Textual Inversion on CustomConcept101 · CustomConcept101 poison average | DINOv2 similarity: 0.21 -> 0.36 | Improved |
| add JPEG preprocessing before loss masking (ablation) | Loss masking defense · CustomConcept101 poison average | DINOv2 similarity: 0.36 -> 0.45 | Improved |
| add JPEG preprocessing before T600 timestep restriction (ablation) | T600 defense · CustomConcept101 poison average | DINOv2 similarity: 0.22 -> 0.43 | Improved |
| combine JPEG, T600 and loss masking (full SZT) versus JPEG+LM (ablation) | SZT · CustomConcept101 poison average | DINOv2 similarity: 0.45 -> 0.46 | Improved |
| apply SZT to poisoned data versus nominal TI on clean data | SZT · CustomConcept101 | DINOv2 similarity: 0.21 -> 0.46 (clean nominal 0.46) | Improved |
| apply PDMPure regeneration defense to clean data | Textual Inversion on CustomConcept101 · CustomConcept101 clean data | DINOv2 similarity: 0.46 -> 0.25 | Hurt |
| raise the high-timestep threshold for TI training on ADM+ data (ablation) | Textual Inversion on NovelConcepts10 · NovelConcepts10 ADM+ | DINOv2 similarity: 0.10 -> 0.21 (t>=600) -> 0.28 (t>=700) | Improved |
| concentrate timesteps at extremely high values (t >= 900) (ablation) | Textual Inversion on NovelConcepts10 · NovelConcepts10 ADM+ | DINOv2 similarity: 0.33 (t>=800) -> 0.26 (t>=900) | Hurt |
| restrict TI training to low timesteps (low threshold) instead of high (ablation) | Textual Inversion on NovelConcepts10 · NovelConcepts10 clean and poisoned | DINOv2 similarity: Clean 0.44 -> 0.17 (low threshold 0.1) | Hurt |
| increase TI learning rate and training steps (ablation) | Textual Inversion on NovelConcepts10 · NovelConcepts10 clean, ADM+, ADM-, EA | DINOv2 similarity | Mixed |
Closest papers by meaning
- From Trojan Horses to Castle Walls: Unveiling Bilateral Data Poisoning Effects in Diffusion Models (NeurIPS 2024 · similarity 0.91)
- Poisoning Attacks on LLMs Require a Near-constant Number of Poison Samples (NeurIPS 2026 · similarity 0.88)
- Attack To Defend: Exploiting Adversarial Attacks for Detecting Poisoned Models (CVPR 2024 · similarity 0.87)
- Intriguing Properties of Diffusion Models: An Empirical Study of the Natural Attack Capability in Text-to-Image Generative Models (CVPR 2024 · similarity 0.87)
- PoisonBench: Assessing Language Model Vulnerability to Poisoned Preference Data (ICML 2025 · similarity 0.86)
- MP-Nav: Enhancing Data Poisoning Attacks against Multimodal Learning (ICML 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.
