Date of Award
Doctor of Philosophy (PhD)
The advent of data mining and machine learning has highlighted the value of large and varied sources of data, while increasing the demand for synthetic data captures the structural and statistical characteristics of the original data without revealing personal or proprietary information contained in the original dataset.
In this dissertation, we use examples from original research to show that, using appropriate models and input parameters, synthetic data that mimics the characteristics of real data can be generated with sufficient rate and quality to address the volume, structural complexity, and statistical variation requirements of research and development of digital information processing systems.
First, we present a progression of research studies using a variety of tools to generate synthetic network traffic patterns, enabling us to observe relationships between network latency and communication pattern benchmarks at all levels of the network stack.
We then present a framework for synthesizing large scale IoT data with complex structural characteristics in a scalable extraction and synthesis framework, and demonstrate the use of generated data in the benchmarking of IoT middleware.
Finally, we detail research on synthetic image generation for deep learning models using 3D modeling. We find that synthetic images can be an effective technique for augmenting limited sets of real training data, and in use cases that benefit from incremental training or model specialization, we find that pretraining on synthetic images provided a usable base model for transfer learning.
Anderson, Jason, "Methods and Applications of Synthetic Data Generation" (2021). All Dissertations. 2917.
Author ORCID Identifier