Generative adversarial network based synthetic data training model for lightweight convolutional neural networks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Inadequate training data is a major hurdle for deep learning, especially with privacy concerns limiting dataset availability.
• Existing methods like data augmentation and transfer learning offer partial solutions but have limitations such as data quality degradation and negative transfer.

Purpose of the Study:

• To propose a novel Generative Adversarial Network-based Synthetic Data Training (GAN-ST) model for generating synthetic data.
• To address the challenge of insufficient training data for training lightweight Convolutional Neural Networks (CNNs).

Main Methods:

• Developed a GAN-ST model integrating enhanced generators based on Deep Convolutional Generative Adversarial Networks (DCGAN) and Conditional Generative Adversarial Networks (CGAN).
• The model employs two independently trained GANs to capture diverse aspects of the original data distribution.
• Evaluated the performance of a CNN classifier trained on both original and GAN-ST-generated synthetic data using MNIST and CIFAR 10 datasets.

Main Results:

• Achieved high classifier accuracy on the MNIST dataset (99.38%) using GAN-ST synthetic data, only 0.05% lower than original data.
• Demonstrated remarkable performance on the CIFAR 10 dataset with 90.23% accuracy using GAN-ST synthetic data.
• Showed significant improvements over single GAN-based training, with gains of 0.66% on MNIST and 7.06% on CIFAR 10.

Conclusions:

• The GAN-ST model effectively generates diverse and realistic synthetic data, improving CNN generalization and classification accuracy.
• This approach offers a viable solution for deep learning applications with limited or sensitive training data.
• The proposed method enhances classifier performance compared to traditional data augmentation and single GAN training strategies.