Automatic Jailbreaking of the Text-to-Image Generative AI Systems
Fuente:
arXiv
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2024
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866917677394558976 |
|---|---|
| author | Kim, Minseon Lee, Hyomin Gong, Boqing Zhang, Huishuai Hwang, Sung Ju |
| author_facet | Kim, Minseon Lee, Hyomin Gong, Boqing Zhang, Huishuai Hwang, Sung Ju |
| contents | Recent AI systems have shown extremely powerful performance, even surpassing human performance, on various tasks such as information retrieval, language generation, and image generation based on large language models (LLMs). At the same time, there are diverse safety risks that can cause the generation of malicious contents by circumventing the alignment in LLMs, which are often referred to as jailbreaking. However, most of the previous works only focused on the text-based jailbreaking in LLMs, and the jailbreaking of the text-to-image (T2I) generation system has been relatively overlooked. In this paper, we first evaluate the safety of the commercial T2I generation systems, such as ChatGPT, Copilot, and Gemini, on copyright infringement with naive prompts. From this empirical study, we find that Copilot and Gemini block only 12% and 17% of the attacks with naive prompts, respectively, while ChatGPT blocks 84% of them. Then, we further propose a stronger automated jailbreaking pipeline for T2I generation systems, which produces prompts that bypass their safety guards. Our automated jailbreaking framework leverages an LLM optimizer to generate prompts to maximize degree of violation from the generated images without any weight updates or gradient computation. Surprisingly, our simple yet effective approach successfully jailbreaks the ChatGPT with 11.0% block rate, making it generate copyrighted contents in 76% of the time. Finally, we explore various defense strategies, such as post-generation filtering and machine unlearning techniques, but found that they were inadequate, which suggests the necessity of stronger defense mechanisms. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2405_16567 |
| institution | arXiv |
| publishDate | 2024 |
| record_format | arxiv |
| spellingShingle | Automatic Jailbreaking of the Text-to-Image Generative AI Systems Kim, Minseon Lee, Hyomin Gong, Boqing Zhang, Huishuai Hwang, Sung Ju Artificial Intelligence Cryptography and Security Recent AI systems have shown extremely powerful performance, even surpassing human performance, on various tasks such as information retrieval, language generation, and image generation based on large language models (LLMs). At the same time, there are diverse safety risks that can cause the generation of malicious contents by circumventing the alignment in LLMs, which are often referred to as jailbreaking. However, most of the previous works only focused on the text-based jailbreaking in LLMs, and the jailbreaking of the text-to-image (T2I) generation system has been relatively overlooked. In this paper, we first evaluate the safety of the commercial T2I generation systems, such as ChatGPT, Copilot, and Gemini, on copyright infringement with naive prompts. From this empirical study, we find that Copilot and Gemini block only 12% and 17% of the attacks with naive prompts, respectively, while ChatGPT blocks 84% of them. Then, we further propose a stronger automated jailbreaking pipeline for T2I generation systems, which produces prompts that bypass their safety guards. Our automated jailbreaking framework leverages an LLM optimizer to generate prompts to maximize degree of violation from the generated images without any weight updates or gradient computation. Surprisingly, our simple yet effective approach successfully jailbreaks the ChatGPT with 11.0% block rate, making it generate copyrighted contents in 76% of the time. Finally, we explore various defense strategies, such as post-generation filtering and machine unlearning techniques, but found that they were inadequate, which suggests the necessity of stronger defense mechanisms. |
| title | Automatic Jailbreaking of the Text-to-Image Generative AI Systems |
| topic | Artificial Intelligence Cryptography and Security |
| url | https://arxiv.org/abs/2405.16567 |