Universal Adversarial Patch Attack for Automatic Checkout Using Perceptual and Attentional Bias.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks (DNNs) are vulnerable to adversarial examples, which are inputs with subtle changes causing misclassification.
• Adversarial patches, localized noise, offer practical attack feasibility but often lack generalization.
• Existing methods fail to generate robust adversarial patches due to ignoring model biases, leading to input-specific attacks.

Purpose of the Study:

• To propose a bias-based framework for generating universal adversarial patches with strong generalization.
• To enhance adversarial patch attack capabilities by exploiting perceptual and attentional biases in DNNs.
• To improve the transferability of adversarial patches across different models and unseen classes.

Main Methods:

• Exploited perceptual bias by extracting textural patch priors from hard examples using style similarities.
• Leveraged attentional bias by confusing model-shared attention patterns for improved transferability.
• Developed a bias-based framework to generate universal adversarial patches.

Main Results:

• The proposed framework generates adversarial patches with strong generalization ability.
• The patches demonstrate improved attacking performance across different classes and models.
• Experiments in digital and physical-world scenarios (Automatic Check-out) show superior performance over state-of-the-art methods.

Conclusions:

• The bias-based framework effectively generates universal adversarial patches with enhanced generalization and transferability.
• Exploiting perceptual and attentional biases is crucial for robust adversarial patch generation.
• The method shows significant potential for real-world applications like Automatic Check-out systems.