Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

Parasite: A Steganography-based Backdoor Attack Framework for Diffusion Models

Created by
  • Haebom

Author

Jiahao Chen, Yu Pan, Yi Du, Chunkai Wu, Lin Wang

Outline

This paper proposes "Parasite," a novel backdoor attack technique for diffusion-model-based image-to-image conversion. While existing backdoor attacks rely on a single, conspicuous trigger to generate a fixed target image, Parasite utilizes steganography to conceal the trigger and incorporates the target content itself as the backdoor trigger, enabling more flexible attacks. Parasite effectively bypasses existing backdoor detection frameworks, and experimental results demonstrate a 0% backdoor detection rate against existing defense frameworks. Furthermore, we analyze the impact of concealment coefficients on attack outcomes through ablation studies.

Takeaways, Limitations

•
Takeaways:
◦
Clearly presenting the backdoor vulnerability of the diffusion model in image-to-image conversion tasks.
◦
Proposal of a new steganography-based backdoor attack technique, "Parasite."
◦
Presents an attack technique with a high success rate that bypasses existing defense techniques.
◦
A new attack method that increases flexibility by utilizing target content as a trigger.
•
Limitations:
◦
Currently, only the attack effectiveness for a specific diffusion model has been verified, and further research is needed to determine the generalizability to other or modified models.
◦
Although it uses steganographic techniques, detection is possible when very sophisticated detection techniques are developed.
◦
Although the attack has a high success rate, further research is needed on its applicability and impact in real systems.
👍
Made with Slashpage