Threat in Frequency: Unveiling Hidden Attack Surfaces by Exploiting Stealthy Backdoors on Diffusion Models

Area of Science:

• Artificial Intelligence
• Machine Learning Security
• Generative Models

Background:

• Diffusion models are advanced generative AI, creating high-quality synthetic data.
• These models are used as end-products (GCAP) and for data augmentation (GCAD).
• Backdoor attacks pose a significant threat to diffusion model security.

Purpose of the Study:

• To address limitations of existing backdoor attacks on diffusion models.
• To propose a novel framework, Evil Diffusion (ELF), for stealthy backdoor attacks.
• To explore vulnerabilities in both GCAP and GCAD scenarios.

Main Methods:

• Introduced ELF-P: a two-stage algorithm for GCAP using frequency-based, imperceptible triggers.
• Developed ELF-D: an algorithm for GCAD that poisons downstream tasks via frequency component perturbations.
• Employed perceptual and defense-resistant constraints in the frequency domain.

Main Results:

• ELF-P successfully created human-imperceptible and defense-resilient triggers for GCAP.
• ELF-D effectively compromised downstream tasks by subtly undermining model performance in GCAD.
• Extensive experiments validated ELF's effectiveness in both attack scenarios.

Conclusions:

• The ELF framework reveals significant hidden vulnerabilities in diffusion models.
• Stealthy backdoor attacks on diffusion models, especially in GCAD, present broader security risks.
• Further research is needed to develop robust defenses against such sophisticated attacks.