# Generating High-fidelity, Synthetic Time Series Datasets with DoppelGANger

**Authors:** Zinan Lin, Alankar Jain, Chen Wang, Giulia Fanti, Vyas Sekar  
**arXiv:** [1909.13403](https://arxiv.org/abs/1909.13403)  
**Date:** 2019-09-30  
**Summary page:** https://www.summarizepaper.com/en/paper/1909.13403v1/

## Key points

- Limited data access is a major challenge in data-driven networking research and development
- Privacy concerns often impede the sharing of confidential information within organizations and with external stakeholders
- Synthetic data models have had limited success due to their narrow scope
- DoppelGANger is a synthetic data generation framework based on generative adversarial networks (GANs)
- DoppelGANger is designed for time series datasets with both continuous and discrete features
- DoppelGANger employs a new conditional architecture that separates metadata generation from time series generation
- DoppelGANger achieves up to 43% better fidelity compared to baseline models
- DoppelGANger captures structural properties of the data that baseline methods are unable to learn
- DoppelGANger provides an easy mechanism for data holders to protect attributes of their data without significant loss of utility
- This research presents a novel approach for generating high-fidelity synthetic time series datasets using GANs, addressing limitations of existing models and offering promising potential for overcoming barriers related to limited data access in networking research and development.

## Summary

Limited data access is a major challenge in data-driven networking research and development. Despite the motivation to share data, privacy concerns often impede the sharing of confidential information within organizations and with external stakeholders such as researchers and vendors. To address this issue, researchers have proposed synthetic data models but these models have had limited success due to their narrow scope. In this study, the authors introduce DoppelGANger, a synthetic data generation framework based on generative adversarial networks (GANs). DoppelGANger is specifically designed for time series datasets that contain both continuous features (e.g., traffic measurements) and discrete features (e.g., protocol name). Modeling time series and mixed-type data has traditionally been difficult but DoppelGANger overcomes these difficulties by employing a new conditional architecture. This architecture separates the generation of metadata from time series while allowing metadata to strongly influence time series generation. The effectiveness of DoppelGANger is demonstrated using three real-world datasets. The results show that DoppelGANger achieves up to 43% better fidelity compared to baseline models and captures structural properties of the data that baseline methods are unable to learn. Additionally, DoppelGANger provides an easy mechanism for data holders to protect attributes of their data without significant loss of utility. Overall, this research presents a novel approach for generating high-fidelity synthetic time series datasets using GANs which addresses the limitations of existing models and provides improved fidelity and structural properties offering promising potential for overcoming barriers related to limited data access in networking research and development.