Frequency-Tuned Universal Adversarial Attacks on Texture Recognition

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Deep neural networks (DNNs) are susceptible to adversarial attacks in natural image classification.
• The impact of adversarial attacks on DNN-based texture recognition remains underexplored.
• Spatial domain perturbation constraints may not effectively limit perceptibility in texture images.

Purpose of the Study:

• To investigate the vulnerability of DNN-based texture recognition to adversarial attacks.
• To propose a novel frequency-tuned universal attack method for texture recognition.
• To evaluate the effectiveness of the proposed method in terms of perturbation perceptibility and attack success rates.

Main Methods:

• Developed a frequency-tuned universal attack method operating in the frequency domain.
• Computed universal perturbations based on local visual frequency characteristics relevant to human perception.
• Evaluated the method on various DNN texture classifiers and datasets using white-box fooling rates.

Main Results:

• The proposed frequency-tuned attack generates less perceptible perturbations compared to existing methods.
• Achieved similar or higher white-box fooling rates on DNN texture classifiers.
• Demonstrated improved attack robustness against defended models and enhanced cross-dataset transferability.

Conclusions:

• Limiting perturbations in the spatial domain is insufficient for texture recognition adversarial attacks.
• The frequency-tuned universal attack is a more effective method for texture recognition.
• The proposed approach enhances the robustness and transferability of adversarial attacks in texture recognition.