Image Representation Induced Subspaces for Practical Classification Robustness.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Image transformations like discrete wavelet transforms (DWTs) and generative models offer meaningful image representations.
• Improving neural network classification robustness against real-world corruptions is a significant challenge.

Purpose of the Study:

• To propose a general method for enhancing neural network classification robustness against real-world corruptions.
• To leverage expressive image representations for improved adversarial robustness.

Main Methods:

• A novel adversarial attack targeting low-dimensional subspaces in transformed image spaces.
• Training neural networks for adversarial robustness using the proposed attack as a proxy for corruption robustness.
• Applying the method with discrete cosine transform (DCT), DWTs, and Glow, focusing on preserving low frequencies or relevant features.

Main Results:

• Models trained with the proposed method show significantly improved robustness against unseen common image perturbations.
• The approach maintains natural accuracy with only a minor sacrifice.
• The method demonstrates generality across different color systems and parameters.

Conclusions:

• The proposed adversarial attack and training strategy effectively enhance neural network robustness against image corruptions.
• Leveraging semantically meaningful image representations in transformed spaces is a promising direction for robust AI.
• The method offers a generalizable solution for improving deep learning model resilience.