Text-to-image (TTI) models offer many innovative services but also raise ethical concerns due to their potential to generate unethical images. Most public TTI services employ safety filters to prevent unintended images. In this work, we introduce the Divide-and-Conquer Attack to circumvent the safety filters of state-of the-art TTI models, including DALL-E 3 and Midjourney. Our attack leverages LLMs as text transformation agents to create adversarial prompts. We design attack helper prompts that effectively guide LLMs to break down an unethical drawing intent into multiple benign descriptions of individual image elements, allowing them to bypass safety filters while still generating unethical images. Because the latent harmful meaning only becomes apparent when all individual elements are drawn together. Our evaluation demonstrates that our attack successfully circumvents multiple strong closed-box safety filters. The comprehensive success rate of DACA bypassing the safety filters of the state-of-the-art TTI engine DALL-E 3 is above 85%, while the success rate for bypassing Midjourney V6 exceeds 75%. Our findings have more severe security implications than methods of manual crafting or iterative TTI model querying due to lower attack barrier, enhanced interpretability , and better adaptation to defense. Our prototype is available at: https://github.com/researchcode001/Divide-and-Conquer-Attack
翻译:暂无翻译