FTA2C: Achieving superior trade-off between accuracy and robustness in adversarial training

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks (DNNs) are vulnerable to adversarial perturbations due to non-robust features.
• Existing adversarial training methods often sacrifice accuracy for robustness by processing features individually.

Purpose of the Study:

• To propose a novel plug-in method, Feature Transformation Alignment and Compression (FTA2C), to enhance DNN robustness and accuracy.
• To introduce a Defense Efficiency Metric (DEM) for evaluating defense strategies.

Main Methods:

• FTA2C employs a compression network to constrain perturbation space and a feature transformation network to enhance robust features.
• An alignment mechanism ensures consistency between adversarial and natural samples in the robust feature space, enabling co-processing.
• The Defense Efficiency Metric (DEM) quantifies the trade-off between natural accuracy and adversarial robustness.

Main Results:

• FTA2C significantly improves adversarial robustness while maintaining high natural accuracy across four benchmark datasets.
• The proposed method demonstrates superior performance in balancing the accuracy-robustness trade-off compared to traditional methods.
• Experiments validate the effectiveness of FTA2C in defending DNNs against adversarial perturbations.

Conclusions:

• FTA2C offers an effective solution for the accuracy-robustness dilemma in adversarial defense.
• The Defense Efficiency Metric (DEM) provides a standardized and interpretable way to evaluate defense methods.
• The proposed approach advances the field of robust machine learning and adversarial defense.