DA3Attacker: A Diffusion-Based Attacker Against Aesthetics-Oriented Black-Box Models.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks (DNNs) excel in image aesthetics assessment (IAA) but lack transparency, hindering commercial use.
• The security and explainability of "black-box" IAA models remain underexplored, creating a significant research gap.
• Understanding model vulnerabilities is crucial for developing robust and trustworthy AI systems in visual analysis.

Purpose of the Study:

• To investigate the susceptibility of current IAA models to adversarial attacks.
• To elucidate the underlying mechanisms contributing to IAA model vulnerabilities.
• To propose a novel framework for generating adversarial examples (AEs) to probe and explain IAA models.

Main Methods:

• A diffusion-based framework, DA3Attacker, utilizing an Attack Diffusion Transformer with aesthetics-oriented filters.
• Two unsupervised training stages to construct a latent space for generating AEs.
• Two controllable attack modes: restricted and unrestricted, for diverse vulnerability testing.

Main Results:

• DA3Attacker effectively deceives diverse black-box IAA models, demonstrating significant vulnerabilities.
• The method provides multi-attribute explanations for feature dependencies in IAA models.
• Extensive experiments on 26 baseline models validate the framework's efficacy.

Conclusions:

• The proposed DA3Attacker framework successfully probes and explains vulnerabilities in IAA models.
• The study contributes evaluation tools, metrics for adversarial robustness, and a dataset of AEs for IAA model enhancement.
• Addressing security and explainability is vital for the broader commercial adoption of IAA technologies.