Learning Deep Features for One-Class Classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• One-class classification is crucial for anomaly and novelty detection.
• Transfer learning offers a way to leverage existing labeled data for new tasks.
• Deep learning models, particularly Convolutional Neural Networks (CNNs), excel at feature extraction.

Purpose of the Study:

• To develop a novel deep-learning-based approach for one-class transfer learning.
• To enable effective feature learning for one-class classification using data from unrelated tasks.
• To improve performance in anomaly detection, novelty detection, and authentication.

Main Methods:

• A deep one-class (DOC) classification method is proposed, operating on a chosen CNN architecture.
• Two novel loss functions, compactness loss and descriptiveness loss, are introduced to optimize feature space.
• A parallel CNN architecture and a template matching framework are utilized for enhanced feature learning and testing.

Main Results:

• The DOC method demonstrates significant improvements over existing state-of-the-art methods.
• Experiments were conducted on publicly available datasets for anomaly detection, novelty detection, and mobile active authentication.
• The approach successfully produces descriptive features with low intra-class variance.

Conclusions:

• The proposed deep one-class transfer learning method is effective for various detection and authentication tasks.
• The novel loss functions and architecture contribute to superior performance in one-class classification.
• This work advances the field of one-class classification by enabling efficient learning from unrelated data.