ANOMALY DETECTION USING GENERATIVE ADVERSARIAL NETWORKS

Abstract

Anomaly detection (AD) has emerged as a critical application across various domains, especially where identifying abnormal behaviors or events is crucial. The advent of deep learning techniques has significantly advanced AD methods, enabling the han dling of complex and high-dimensional data. However, these advancements pose the challenge of explainability, requiring approaches that address the ’black box’ nature of deep learning models. This thesis builds upon a comprehensive review of recent AD techniques, emphasizing their explainability within the realm of Explainable AI (XAI). Key insights include the importance of interpretability in AD systems, the versatility of deep learning architectures, and emerging trends such as graph-based AD using deep learning. Building on this theoretical foundation, the thesis also explores practical enhance ments through the implementation of the Skip-GANomaly model with novel modifica tions to its loss function, incorporating contrastive learning to improve semi-supervised AD. Contrastive learning involves training a model to distinguish between positive and negative sample pairs, leading to robust representation learning. Experimental results demonstrate that these modifications yield significant performance improve ments across various datasets. By integrating a thorough exploration of XAI in AD and proposing an effective semi-supervised AD approach, this thesis aims to advance the field, providing valuable insights and paving the way for future research and appli cations.