You do not have permission to edit this page, for the following reason:
The action you have requested is limited to users in the group: Users.
Project description (free text)
Give a concise project description. Include:
Access to high-quality big datasets for improving deep learning (DL)-based models is a big challenge, more specifically in overly sensitive applications such as healthcare systems where maintaining data privacy is a necessity. Synthetic data generation is a principal tool for various users from researchers who leverage data for models’ training, to educators who aim to teach statistical approaches. The aim of using synthetic data can be categorized into several essential groups such as protecting privacy. However, for generating synthetic data using DL models (like generative adversarial networks (GANs)) we need the training data, and it has been proved that the gradient parameters of these models can remember the training data. For focusing on the privacy-preserving issue of training data in synthetic health data generation, we are going to modify the idea of privacy-preserved GANs [1] to suitable GANs for time series data [2]. Time-series GANs are applicable for generating synthetic electrical health records (EHRs) [3]. In this master thesis, we aim to study different differential privacy-preserving methods to add well-designed noise to the gradients during the training phase of time-series GANs. Conclusion: In this master thesis, we aim to study privacy-preserving approaches in deep learning and develop a model that preserves the privacy of training data in the processes of generating synthetic data. The title of this thesis in detail is ‘the privacy-preserving aspect of generating synthetic electrical health records (Synthetic-EHRs) using time-series generative adversarial networks (Time-GANs).
Summary:
This is a minor edit Watch this page
Cancel
Home
Research
Education
Partners
People
Contact
