Digital Well-Being

Inspiration

In today's world, dependence on digital sources of information has become very common. Unauthorized distribution of user data such as Facebook data leaks, hacking of social network accounts, misuse of deep learning techniques like Generative Adversal Networks to create fake photos, leverage of social media to spread fake news and controversy, adds chaos and confusion to a user's digital experience. I thought of combating the evil of this era by using deep learning to help millions of internet users in getting some peace.

What it does

We have built a series of Digital Wellbeing tools that are aimed to help people detect and identify fake news, by analyzing the news article through state of the state of the art Hybrid CNN and Bi-Directional LSTM Models. Also to accompany the fake news analyzer we have also built fake image detector which detect image modification, image morphing or manipulation to help people fight against rising face image manipulation.

How we built it

The fake news analyzer is based on (Fake News Identification on Twitter with Hybrid CNN and RNN Models, by Oluwaseun Ajao, Deepayan Bhowmik, Shahrzad Zargari), and is built by using FastAI's library - Pytorch in python. The Dataset the used to train is FNC1 and few web articles scraped by us. The image modification detect is completely based on (Image Splice Detection, by Minyoung Huh) and is written in Tensorflow and Keras. The frontend is completely based on Flask-python (REST API) and HTML javascript and CSS.

Challenges we ran into

Implementing Image Splice Detection was extremely difficult and the data set that would train our model as the pertained model required us to have GPU access. With no access to GPU and Limited data set model was still be achieved a respectable score of 88% accuracy for splice detection can be improved (Paper has 95% accuracy) and 81.6% accuracy for Fake news detector.

Accomplishments that we're proud of

We build it!!!!!!!!!!!

What we learned

Deep learning, How key role does dataset plays in training a model and how much effort goes to deciphering research paper :)

What's next for Digital Well-Being

We really wanted to build more and integrate our chrome extension with facebook and twitter but we were impeded by the lack of time. We would love to continue working and developing this project further as we truly believe in empowering users and using our knowledge to help them.

Built With

• bigru
• cnn
• deeplearning
• fastai
• flask
• google-cloud
• html5
• javascript
• keras
• lstm
• python
• pytorch
• tensorflow