Internet of Things Anomaly Detection Enhancement using GAN-based Data Augmentation

Anomaly detection in Internet of Things (IoT) systems is essential for preventing failures, security breaches, and inaccurate predictions. Deep learning-based approaches are effective for this purpose but often suffer from insufficient labeled training data, which limits detection accuracy. To address this challenge, we propose a novel anomaly detection framework that embeds synthetic data generation with the anomaly detection process. The framework integrates a Generative Adversarial Network (GAN)-based module that learns underlying patterns in normal and/or abnormal data to generate synthetic samples, effectively augmenting the training set. This enriched dataset allows the anomaly detection module to better capture data variability and improve accuracy. For anomaly detection, raw data is first converted into spectrogram images to enhance feature representation, which are then encoded by a pre-trained Convolutional Neural Network (CNN) for effective feature extraction. These features are passed to one of two detection models: a deep Autoencoder, which learns representations of normal data in an unsupervised manner and identifies anomalies by reconstruction error, or a CNN classifier, which uses supervised learning to classify normal and abnormal data. Experimental results show that GAN-based data augmentation significantly boosts detection performance in both the supervised and unsupervised learning cases, making this approach highly effective for reliable anomaly detection in IoT systems.