EEG Signal Generation

TLDR;

• Building an EEG generation and anonymization model working with Josh Sanz, Ryan Kaveh, as well as Professors Anant Sahai and Rikky Muller
• Fine-tuning stable diffusion model in order to be able to generate unique EEG spectrograms that retains pertinent information while removing features that can be used in order to determine gender and age
• Use ear EEG classification transformer model as an adversarial network to determine if generated data is convincing

Motivation

EEG signals are highly unique, there have been papers published showing that in some cases they can be used similiar to a finger print. This makes many people hesitant to make publicly avaliable, as if one can figure out what hospital and when they were taken you can use public data to trace who the patient is. If one were to be able to randomize certain parts of a dataset while retaining other important information this would allow for datasets to be anonymized.

The Problem Statement

If one were to want to modify EEG signals they would have to do a series of things: take in EEG signals through some medium, interpret signal to classify certain features, and be able to produce new EEGs with those features altered.

Implementation

We have completed the investigation on this project, and are now actively writing this.

I will give a full breakdown when a paper has been publicly published.

Last updated: 6/15/2024