A comprehensive Approach with Generative Based Augmentation and Deep Learning to Enhance Anomaly Detection in Limited Dataset

Across numerous study fields, anomaly detection is a challenging issue to solve. Unseen anomalies indicate a complicated problem that has generated a variety of solutions throughout the years: correctly identifying and classifying them. Overfitting has a major impact on traditional tree-based anomaly detection methods. Recently, adversarial training and generative adversarial networks (GANs) have emerged as powerful techniques for addressing this problem and yielding remarkable outcomes. By combining GAN-based data augmentation, and model designing through Convolutional Neural Network (CNN) models, and transfer learning demonstrates an original method to reduce overfitting in anomaly detection in comparison with conventional tree based techniques. By avoiding 100% accuracy, a classic indication of overfitting, anomaly detection using this procedure computed the best accuracy. The work further explores transfer learning with some pre-trained models that have undergone significant training on massive image datasets. In conclusion, a comparison of the Decision Tree, CNN, and Transfer Learning using ResNET before and after data augmentation enables the selection of the model that best balances accuracy and overfitting, thereby tackling the complex problem of anomaly detection.