Abstract
The problem of data skewness, class imbalance, data scarcity and noise limits the application of machine learning and deep
learning models in applications like anomaly detection, fraud detection, intrusion detection, fault diagnosis, machine-tomachine communication, etc. Performance of supervised learning leans towards majority class and fails to generalize testing
data in class imbalance and noisy data problems. Using neural-based data augmentation techniques for data generation and
deep convolutional models for classification would enhance the performance of the applications mentioned above. Recently,
GANs (generative adversarial networks) showed significant improvements in generating images. In this paper, a model that
uses a conditioned deep convolutional GAN and an auxiliary classifier are proposed to tackle the aforementioned issues.
The conditioned GAN is used for data generation of minority classes images and noisy images. Another auxiliary deep
convolutional model is employed for the classification of images on data augmented dataset. Also, a multi-hinge loss is
employed in both the data augmentation and classification tasks. The effectiveness of the proposed model is investigated on
the quality of generated images and classification metrics using four publicly available popular datasets: MNIST, EMNIST,
CIFAR10, and SVHN. The proposed model has shown significant improvements over the other often used models of data
augmentation and multi-class imbalance image classification in terms of generated samples and classification accuracy.
Keywords Imbalanced data · Convolutional neural network · Data augmentation · Generative model · Adversarial network