Multichannel One-Dimensional Data Augmentation with Generative Adversarial Network.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Data augmentation is crucial in deep learning, with existing methods primarily targeting computer vision.
• One-dimensional data augmentation, particularly for multichannel data, remains an underexplored area.
• Generative Adversarial Networks (GANs) are powerful tools for data generation but require adaptation for specific data types.

Purpose of the Study:

• To propose a GAN-based data augmentation technique for generating multichannel one-dimensional data from single-channel inputs.
• To address the limitations of existing data augmentation methods in the context of one-dimensional datasets.
• To enhance the performance of deep learning models by providing diverse and realistic synthetic data.

Main Methods:

• Developed a GAN architecture incorporating multiple discriminators, adapting Deep Convolutional GAN (DCGAN) and PatchGAN.
• The architecture is designed to capture both global patterns and local channel-specific information in the generated multichannel data.
• Utilized website fingerprinting data for experimental validation.

Main Results:

• The proposed GAN-based data augmentation model achieved significantly improved results compared to vanilla GAN.
• Achieved low Fréchet Inception Distance (FID) scores of 0.005, 0.017, and 0.051 for the three data channels, indicating high-quality data generation.
• Vanilla GAN resulted in substantially higher FID scores (0.458, 0.551, 0.521), demonstrating the effectiveness of the proposed approach.

Conclusions:

• The proposed GAN-based method is effective for data augmentation of multichannel one-dimensional data.
• The architecture's ability to leverage multiple discriminators enhances the quality and relevance of generated synthetic data.
• This work opens new avenues for data augmentation in domains dealing with one-dimensional time-series or sequential data.