Dormant key: Unlocking universal adversarial control in text-to-image models.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Cybersecurity

Background:

• Text-to-Image (T2I) diffusion models excel at image generation but pose security risks.
• Malicious prompt modifications can bypass safety filters, generating misleading or Not-Safe-For-Work (NSFW) content.
• Existing adversarial attacks lack generalizability and are detectable by current defenses.

Purpose of the Study:

• To propose a universal adversarial attack framework for T2I diffusion models.
• To develop a method that bypasses existing safety mechanisms and improves attack transferability and imperceptibility.
• To address the limitations of prompt-specific attacks and detectable text-space perturbations.

Main Methods:

• Introduced 'dormant key,' a universal adversarial attack framework.
• Developed a transferable suffix that acts as a plug-in for any text input.
• Implemented a hierarchical gradient aggregation strategy for robust optimization across diverse prompts.

Main Results:

• The 'dormant key' framework demonstrated effective balance between attack performance and stealth.
• Achieved over 18% improvement in success rate for NSFW generation tasks compared to baselines.
• Successfully bypassed major safety mechanisms including keyword filtering, semantic analysis, and text classifiers.

Conclusions:

• The proposed 'dormant key' framework offers a robust and transferable adversarial attack for T2I models.
• This method enhances the security risks associated with T2I diffusion models, particularly for NSFW content generation.
• The findings highlight the need for advanced defense mechanisms against sophisticated adversarial attacks in AI image generation.