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 LeeICCV 2025

TL;DR

This paper introduces ZIUM, a novel adversarial attack method that can generate images containing "unlearned" concepts from diffusion models, even after those concepts have been removed. ZIUM excels at customizing these generated images to match an attacker's specific intent and uniquely enables "zero-shot" attacks, meaning it can reuse learned attack parameters for new images without re-optimization, significantly reducing attack time.

What Problem Does This Paper Address?

Machine unlearning (MU) is a technique used to remove specific data points or concepts from deep learning models, often to enhance privacy or prevent the generation of undesirable content (like NSFW material). However, even after unlearning, these models can still be tricked into generating the removed concepts using adversarial prompts. Existing methods for these attacks often struggle to:

  1. Align with attacker intent: They can generate the unlearned concept but may not incorporate specific stylistic or contextual preferences of the attacker.
  2. Be computationally efficient: Identifying effective adversarial prompts can be a costly, iterative process, especially when targeting new variations of an unlearned concept.
  3. Perform "zero-shot" attacks: Most methods require re-optimization for each new target image or concept, making repeated attacks time-consuming.

This paper addresses these limitations by proposing ZIUM, a method that allows for intent-aware adversarial attacks with zero-shot capabilities.

Methods

ZIUM operates in two phases: an Initial Attack Phase and a Zero-Shot Attack Phase.

Initial Attack Phase: This phase focuses on learning the optimal adversarial condition. It involves:

  • Visual-Text Alignment: ZIUM extracts visual features from a "target attack image" (which contains the unlearned concept) using a visual encoder. These features are then transformed into text embeddings using an image captioning module and a projection layer. This process aims to capture the essence of the unlearned concept in a text format.
  • Optimization Process: This phase uses a diffusion classifier mechanism to find the best adversarial condition. It leverages both the text embedding derived from the target attack image and a "user-intent prompt" provided by the attacker. The goal is to maximize the probability of generating an image that incorporates both the unlearned concept and the attacker's intent. This optimization is performed by updating the parameters of the image captioning module and projection layer, while keeping the visual encoder fixed to reduce computational cost and prevent overfitting.

Zero-Shot Attack Phase: This is the core innovation for efficiency. Once the initial attack phase has optimized the visual-text alignment modules, ZIUM can perform attacks without further optimization.

  • Reusing Learned Parameters: The optimized image captioning module and projection layer are reused.
  • Flexible Input: Attackers can now use unseen target attack images and modify user-intent prompts freely. ZIUM uses these new inputs with the pre-learned alignment to generate images that reflect the unlearned concept and the new attacker intent. This eliminates the need for costly re-optimization for each new attack scenario or target image.

Datasets

The paper uses several datasets for evaluating different unlearned concept scenarios:

  • Nudity: NSFW dataset (142 prompts).
  • Violence & Illegal Activity: I2P dataset (selected 334 prompts for violence, 248 for illegal activity).
  • Style: Van Gogh's artistic style (50 prompts).
  • Object: Church and parachute (50 prompts each).

Target attack images were generated using the vanilla Stable Diffusion 1.4v model based on prompts from these datasets.

Key Insights and Ideas

  • Intent Matters: The paper highlights that simply generating an unlearned concept isn't enough; aligning the output with the attacker's specific intent (e.g., "Japanese comics" style, "at the beach" setting) is crucial for realistic adversarial attacks.
  • Visual-Text Bridge: Transforming visual embeddings of the target attack image into text embeddings is key to bridging the gap between the visual unlearned concept and the text-based user-intent prompt.
  • Zero-Shot Efficiency: The ability to reuse learned parameters for subsequent attacks on the same unlearned concept is a significant breakthrough, drastically reducing attack time and computational overhead.
  • Target Attack Image Superiority: The results suggest that using a target attack image that explicitly contains the unlearned concept is more effective than relying solely on text prompts for generating adversarial content.

Key Results

  • Superior Attack Success Rate (ASR): ZIUM outperforms existing adversarial attack methods by at least 22.6%p and up to 62.0%p on average across various unlearned concept scenarios and models.
  • Effective Intent Customization: ZIUM successfully generates images that accurately reflect attacker intents, as demonstrated by examples showing changes in background, objects, and styles based on user-intent prompts.
  • Significant Time Reduction (Zero-Shot): ZIUM's zero-shot attack reduces attack time by at least 45.5 times and up to 149.5 times on average compared to existing methods that require re-optimization for each attack.
  • Consistent Performance: ZIUM shows relatively consistent performance across different scenarios, with a variation of only 39.4%p, unlike other methods that exhibit larger performance swings.

Limitations and Future Work

  • Limitations:
    • While ZIUM achieved superior ASR in most scenarios, it did not achieve a superior ASR compared to existing methods for "illegal activity," "church," and "parachute" categories in some instances, though the difference was often negligible.
  • The paper notes that for "illegal activity," "church," and "parachute" categories, ZIUM did not achieve a superior ASR compared to existing methods in some cases.
  • Future Work:
    • Develop a prompt engineering mechanism to automate the transformation of given conditions into text prompts.
  • Create a model-agnostic mechanism by applying transferable adversarial attack methods.

Summary updated

BibTeX
@inproceedings{yook2025zium,
  author = {Hyun Jun Yook and Ga San Jhun and Jae Hyun Cho and Min Jeon and Donghyun Kim and Tae Hyung Kim and Youn Kyu Lee},
  title = {{ZIUM: Zero-Shot Intent-Aware Adversarial Attack on Unlearned Models}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2025},
}
Figure