Attack to Fool and Explain Deep Networks.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep visual models are vulnerable to adversarial perturbations, often perceived as noise by humans.
• This vulnerability has fueled debate about the alignment between deep visual representations and human perception.
• Existing research often overlooks the potential for human-interpretable structures within adversarial examples.

Purpose of the Study:

• To challenge the view that deep visual representations are misaligned with human perception.
• To propose a novel adversarial attack that generates human-meaningful patterns.
• To demonstrate the utility of these perturbations for explaining deep visual models and enabling new applications.

Main Methods:

• Developed a targeted adversarial attack to confuse object categories while respecting semantic boundaries.
• Analyzed the geometric patterns and information revealed by the generated perturbations.
• Adapted the adversarial objective to create an 'explainability' tool for deep visual representations.

Main Results:

• The proposed attack generates regular geometric patterns in perturbations, indicating structured information.
• These perturbations provide insights into the decision boundaries of deep visual models.
• The method successfully visualizes model understanding of human-defined semantic notions.

Conclusions:

• Adversarial perturbations can contain human-meaningful patterns, suggesting a degree of alignment with human perception.
• The novel adversarial attack serves as a powerful tool for interpreting and explaining deep visual models.
• The derived perturbations have practical applications in image generation, inpainting, and interactive manipulation.