Toward Transferable Defense Against Malicious Image Edits.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Cybersecurity

Background:

• Diffusion models enable advanced image editing but are vulnerable to malicious manipulations.
• Existing defenses lack transferability across different editing models.
• Imperceptible perturbations show potential but face limitations.

Purpose of the Study:

• To propose a novel framework, Transferable Defense Against Malicious Image Edits (TDAE), for robust defense against malicious image edits.
• To enhance image immunity against unseen editing models and achieve cross-model transferability.
• To improve the security of diffusion-based image editing systems.

Main Methods:

• Introduced a bimodal framework combining visual and textual defense strategies.
• Developed the FlatGrad Defense Mechanism (FDM) for gradient regularization, promoting robustness at flat minima.
• Proposed Dynamic Prompt Defense (DPD) using adversarial optimization of text embeddings for cross-model alignment.

Main Results:

• TDAE demonstrated state-of-the-art performance in mitigating malicious edits.
• Achieved significant robustness improvements under both intra-model and cross-model evaluations.
• FDM and DPD mechanisms effectively enhanced image immunity and cross-model transferability.

Conclusions:

• The proposed TDAE framework offers a significant advancement in defending diffusion-based image editing systems.
• Coordinated image-text optimization is crucial for achieving transferable defense against malicious edits.
• TDAE provides a robust solution for enhancing image security in generative AI applications.