M3C: Resist Agnostic Attacks by Mitigating Consistent Class Confusion Prior.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep neural networks (DNNs) face significant security challenges from adversarial attacks.
• Adversarial Training (AT) is effective but often lacks generalizability against unseen attacks.
• Existing AT methods struggle with training-agnostic attacks due to limited robustness.

Purpose of the Study:

• To address the limited generalizability of current adversarial training (AT) methods.
• To propose a unified defense strategy against a wide range of adversarial attacks, including training-agnostic ones.
• To introduce a novel approach that leverages a consistent class confusion (3C) prior for enhanced robustness.

Main Methods:

• Identified a generalizable prior: consistent class confusion (3C) in AT classifiers across diverse attacks.
• Proposed a Mitigating Consistent Class Confusion (M3C) model to enhance AT generalizability.
• Optimized an Adversarial Confusion Loss (ACL) weighted by uncertainty to focus on confused samples.
• Introduced a Gradient-Aware Attention (GAA) mechanism to suppress malignant features and enhance correct class confidence.

Main Results:

• The M3C model significantly improves the generalization of AT robustness against agnostic attacks.
• Experiments on multiple benchmarks and network frameworks validate the effectiveness of the proposed method.
• Demonstrated that mitigating consistent class confusion leads to more robust DNNs.

Conclusions:

• The consistent class confusion (3C) prior offers a unified perspective for defending against diverse adversarial attacks.
• The M3C approach provides a promising direction for developing more generalizable and robust adversarial defense strategies.
• This research opens new avenues for overcoming the challenge of training-agnostic attacks in deep learning.