Toward Intrinsic Adversarial Robustness Through Probabilistic Training.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning Security

Background:

• Deep neural networks (DNNs) are vulnerable to adversarial perturbations, posing security risks.
• Current adversarial training methods optimize natural examples, ignoring adversarial domain adversaries, leading to suboptimal decision boundaries and reduced robustness.

Purpose of the Study:

• To propose Adversarial Probabilistic Training (APT) to enhance DNN robustness.
• To bridge the distribution gap between natural and adversarial examples by modeling latent adversarial distributions.

Main Methods:

• Estimating adversarial distribution parameters at the feature level for efficiency.
• Decoupling distribution alignment using an adversarial probability model and original adversarial examples.
• Implementing a novel reweighting mechanism considering adversarial strength and domain uncertainty.

Main Results:

• APT demonstrates superior performance against diverse adversarial attacks.
• The method shows effectiveness across different datasets and scenarios.
• The proposed approach mitigates the negative impact of optimization bias.

Conclusions:

• Adversarial Probabilistic Training (APT) offers a more effective strategy for improving adversarial robustness in DNNs.
• The feature-level estimation and novel reweighting mechanism contribute to enhanced security against sophisticated attacks.