Comparative analysis and Evaluation of techniques for Generating High-Quality synthetic Datasets for Industrial Control Systems

Abstract

Industrial Control Systems (ICSs) and SCADA networks are vital for managing complex industrial infrastructures, ensuring smooth operations across applications. Rising cyber threats prompt the exploration of machine learning and deep learning techniques, utilizing neural networks to detect and predict attacks. However, limited training data and biased outcomes undermine these models' accuracy. Privacy concerns add complexity.

Synthetic data generation emerges as a research focus. The goal is to replicate real data's statistical features for augmentation, privacy, and model development. Balancing realism and confidentiality is crucial. Evaluating synthetic data is challenging. Existing methods cater to specific applications, demanding an unbiased, diverse, standardized evaluation.

This thesis performs a comprehensive comparative analysis of synthetic data generation for ICS datasets. It proposes an evaluation framework using visualization and statistics. Three models—GANs, VAEs, GMMs—are compared, assessing Fidelity, Privacy, Diversity, Interpretability, and Utility. The aim is to guide researchers and practitioners in method selection for ICS applications, promoting diverse, unbiased datasets.

The analysis highlights the strengths, limitations, and trade-offs of synthetic data techniques for ICS datasets. Findings aid optimal high-quality synthetic data generation, enabling privacy-preserving research. Diverse synthetic datasets facilitate experimentation, and validation, bolstering ICS robustness. This research advances ICS understanding, fostering secure and efficient development.