Differentially Diagnosed
This is my most techincal project to date. I created it for my CS3110 class Data Privacy’s final project. It is the longest I have spent on the project and I am still continously working on it trying to obtain better accuracy and more in-depth analysis. The jupyter notebook in the github repository has more explanations, but here I will give a brief synopsis along with some analysis.
Why Did I Make This?
For the Data Privacy final project, us students had many different options to choose from. I could have submitted an entry to Open-DP’s real world implementations of differential privacy, analyzed a current deployment of it, or take a new dataset and implement different variants of differential privacy. But as I progressed through this class I began to notice that I was doing further reading after every class and became very intrigued by differential privacy. It was at that moment I decided to go above and beyond for this project.
How It Worked (Initially)
The original neural network that I submitted for my final project is three layers and it acheived an accuracy a little above 90%. I trained it using stochastic gradient descent with a learning rate of 0.1, a momentuam value of 0.9, and 50 epochs. The data set contains 100,000 rows and my mini-batch size was 64 with shuffling enabled.
The highest epsilon value () I tested was 9.30 and the lowest was 0.33. With the higher epsilon value, the model accuracy changed by a little more than 1%, and the lower accuracy dropped >10%. Below I have the graph showing the differences in accuracy compared to epsilon values. (Need to insert this)
Current Structure
I am currently reworking the project now that the semester is over to be a four layer neural network with a mini batch size of 265. I also added a dropout rate of 0.2 which has helped with the acccuracy. I am going to analyze the affects of different noise multipliers thoroughly and attached to this will be a nice paper by the end of winter break so stay tuned.
Further Ideas
Something I’ve been thinking about is adding another level to this project. From the initial dataset, I dropped the column diabetes stage. What I am thinking about now is having another target column determining a patients diabetes stage based on the original features, and if they are diagnosed.
Look at the Code
Here is the link to the repository on github. Repository