Advancing Adversarial Training by Injecting Booster Signal.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep neural networks (DNNs) are vulnerable to adversarial attacks.
• Adversarial training (AT) is effective but can reduce natural accuracy.
• Existing methods focus on optimizing model parameters for defense.

Purpose of the Study:

• To propose a new method for improving adversarial robustness and natural accuracy in DNNs.
• To introduce an external signal, termed a booster signal, as a defense mechanism.
• To enhance existing adversarial training methods without altering model parameters.

Main Methods:

• An external booster signal is injected into images, separate from original content.
• The booster signal is optimized collaboratively and in parallel with model parameters.
• This approach aims to boost both adversarial robustness and natural accuracy.

Main Results:

• The proposed booster signal significantly improves both natural and robust accuracies.
• The method demonstrates superior performance compared to state-of-the-art adversarial training techniques.
• The booster signal approach is general and adaptable to various AT methods.

Conclusions:

• The booster signal offers an effective, parameter-free approach to enhance DNN security.
• This method successfully addresses the trade-off between natural accuracy and adversarial robustness.
• The flexibility of the booster signal allows for broad application in defending DNNs against attacks.