Research map / Adversarial robustness and security / Attacks on text and graphs
ZIUM: Zero-Shot Intent-Aware Adversarial Attack on Unlearned Models
Hyun Jun Yook, Ga San Jhun, Jae Hyun Cho, Min Jeon, Donghyun Kim, Tae Hyung Kim, Youn Kyu Lee
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 |
|---|---|---|---|
| condition unlearned diffusion model on the target attack image via image-captioning-based visual-text alignment | ZIUM · Unlearned models, average over scenarios | attack success rate (ASR): 65.2% -> 87.8% (UnlearnDiffAtk vs ZIUM) | Improved |
| use ZIUM on illegal activity, church and parachute scenarios | ZIUM · Illegal activity, church, parachute | attack success rate (ASR): e.g. church ESD 66.0% (UnlearnDiffAtk) vs 62.0% (ZIUM) | No clear effect |
| use target attack images instead of target prompts | attack methods · Unlearned models, average over scenarios | attack success rate (ASR): 65.2% and 87.8% (image-based) vs 25.8% and 30.2% (prompt-based) | Improved |
| reuse the optimized captioning module and projection layer for zero-shot attack instead of re-optimizing (ablation) | ZIUM on ESD nudity · Nudity, ESD | attack success rate (ASR): 97.1% -> 84.5% | Hurt |
| use zero-shot attack phase rather than per-attack optimization (ablation) | ZIUM on ESD · ESD attack time | attack time (mins): 9.0 -> 0.2 | Improved |
| add user-intent prompt concatenated to the aligned image embedding | ZIUM · qualitative user-intent customization | alignment with attacker intent | Improved |
| apply ZIUM instead of baselines against AdvUnlearn on nudity | AdvUnlearn nudity · Nudity, AdvUnlearn | attack success rate (ASR): 21.1% -> 91.5% | Improved |
Closest papers by meaning
- Rethinking Adversarial Robustness in the Context of the Right to be Forgotten (ICML 2024 · similarity 0.85)
- Flexible, Efficient, and Stable Adversarial Attacks on Machine Unlearning (ICML 2025 · similarity 0.85)
- Unlearn and Burn: Adversarial Machine Unlearning Requests Destroy Model Accuracy (ICLR 2025 · similarity 0.84)
- Defensive Unlearning with Adversarial Training for Robust Concept Erasure in Diffusion Models (NeurIPS 2024 · similarity 0.83)
- Generate Universal Adversarial Perturbations for Few-Shot Learning (NeurIPS 2024 · similarity 0.83)
- Towards LLM Unlearning Resilient to Relearning Attacks: A Sharpness-Aware Minimization Perspective and Beyond (ICML 2025 · similarity 0.82)
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.
