Shaping Deep Feature Space Towards Gaussian Mixture for Visual Classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks commonly use softmax cross-entropy loss for training.
• Existing methods may lack robustness against adversarial attacks and precise feature space modeling.

Purpose of the Study:

• To propose a novel Gaussian Mixture (GM) loss function for deep neural networks in visual classification.
• To enhance classification performance and adversarial robustness by explicitly shaping the feature space.

Main Methods:

• Developed a Gaussian Mixture (GM) loss function that shapes deep feature space towards a Gaussian Mixture distribution.
• Incorporated a classification margin and likelihood regularization into the GM loss.
• Utilized the discrepancy between clean and adversarial feature distributions for adversarial example detection.

Main Results:

• The GM loss facilitates high classification performance and accurate feature distribution modeling.
• The GM loss effectively distinguishes adversarial examples.
• Theoretical analysis shows the GM loss achieves a symmetric feature space, enhancing adversarial robustness.
• The proposed method shows favorable performance on image classification, face recognition, and detection tasks.

Conclusions:

• The Gaussian Mixture loss function offers improved classification accuracy and adversarial robustness.
• It provides a principled way to model feature distributions and defend against adversarial attacks.
• The GM loss is efficient, requires no additional trainable parameters, and is broadly applicable.