S⁴ST: A Strong, Self-Transferable, faSt, and Simple Scale Transformation for Data-Free Transferable Targeted Attack

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning Security

Background:

• Transferable Targeted Attacks (TTAs) suffer from overfitting to surrogate models.
• Existing data-free solutions often require black-box feedback, compromising fairness.
• Current methods struggle with strict black-box transfer settings.

Purpose of the Study:

• To propose blind estimation measures for analyzing transformation effectiveness under black-box constraints.
• To challenge conventional assumptions in TTA research.
• To develop a data-free TTA method with state-of-the-art effectiveness and efficiency.

Main Methods:

• Introduced self-alignment and self-transferability measures for transformation analysis.
• Investigated the impact of basic image transformations (scaling, geometric, color) on TTA.
• Designed S⁴ST (Strong, Self-transferable, faSt, Simple Scale Transformation) integrating scaling, low-redundancy, and block-wise operations.

Main Results:

• Simple scaling transformations uniquely enhance targeted transferability.
• Geometric and color transformations show high internal redundancy.
• S⁴ST achieves state-of-the-art effectiveness-efficiency balance without data dependency across diverse settings.

Conclusions:

• Scaling's effectiveness is linked to the multi-scale nature of visual data and training augmentations.
• The proposed blind estimation measures and S⁴ST framework offer a robust, data-free solution for TTAs.
• The method demonstrates strong generalization capabilities on medical imaging and face verification tasks.