Overview

This project is a data-driven approach to detecting Russian Disinformation utilizing a dataset of Internet Research Agency ads released by the U.S. House Intelligence Committee as part of an investigation into the alleged Russian Disinformation Campaign in the 2016 U.S. Election.

This project is carried out by Faculty members Dr. Farnoush Banaei-Kashani and Dr. Haadi Jafarian as well as student Tobby Lie at the Computer Science school of the University of Colorado Denver

Methods

Machine Learning Models on Textual Data

We utilized Support Vector Machine and Naive Bayes based solutions to training on textual data in the dataset. We gained successful metrics from this which can be viewed in our plots directory.

Deep Convolutional Neural Networks on Image Data

This is an ongoing area we are experimenting with. We have gained successful metrics via VGG16 and ResNet50. These results can also be viewed in the Russian-Disinformation-Project/CNN_experiments/CNN experiments metrics directory.

Latent Dirichlet Allocation to Topic Model

We utilized LDA to derive the dominant themes in our data and to draw correlations between the themes and data samples themselves.

Future Work

We intend to co-train our models to effectively combine our efforts in image and textual data training. This will be carried out after our CNN training has been finalized.