Related Experiment Video
Updated: May 5, 2026
Advanced Diffusion Imaging in The Hippocampus of Rats with Mild Traumatic Brain Injury
Published on: August 14, 2019
Adversarial discriminant attack on text-to-image diffusion models
Hanxiao Wu1, Shengwu Xiong2, Dong Yi3
1School of Computer Science and Artificial Intelligence, Wuhan University of Technology, Wuhan, 430070, Hubei, China; Institute of Automation, Chinese Academy of Sciences, Beijing, 100190, China; School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, 101408, China; Wuhan AI Research, Wuhan, 430000, Hubei, China.
A new attack method, Adversarial Discriminant Attack (ADAtk), effectively bypasses safety mechanisms in concept-erased diffusion models. ADAtk generates images classified as Not-Safe-For-Work (NSFW) with over 90% success, revealing vulnerabilities in current AI safety techniques.
Area of Science:
- Artificial Intelligence
- Computer Vision
- Generative Models
Background:
- Concept-erased diffusion models face challenges in preventing Not-Safe-For-Work (NSFW) content generation.
- Existing attack methods focus on image similarity, which does not guarantee successful NSFW reconstruction.
Purpose of the Study:
- To propose a novel attack method, Adversarial Discriminant Attack (ADAtk), to expose vulnerabilities in concept-erased diffusion models.
- To address the limitations of existing generation-focused attacks by adopting a discriminative approach.
Main Methods:
- ADAtk optimizes the likelihood of generating NSFW content by creating adversarial perturbations in the model's latent space.
- The method guides image reconstruction towards a target discriminant class, aiming for classification as inappropriate.
Main Results:
- ADAtk achieved over a 90% success rate in bypassing current internal security mechanisms.
- The attack successfully exposed critical limitations in existing concept-erasure techniques for diffusion models.
Conclusions:
- ADAtk provides crucial insights into improving the safety and reliability of text-to-image generation systems.
- The findings pave the way for developing more secure generative AI models and robust safety protocols.
