Master's Project - Video Manipulation Detection via Recurrent Residual Feature Learning Networks

Link to GitHub: https://github.com/alswilli/Video-Manipulation-Detection-via-Recurrent-Residual-Feature-Learning-Networks

Link to Publication: https://ieeexplore.ieee.org/document/8969458

Visualization of the final neural network architecture, with corresponding layer input shapes listed below each layer block and sizes listed below each layer label

Summary:

For my Master's Project at UC Santa Cruz, I created a novel prototype for a Deep Learning architecture that utilizes Recurrent and Convolutional Neural Networks to enable detection of object insertion, compression, black out, and blurring in videos on a frame-to-frame basis. I worked on this project with one other M.S./PhD student under the supervision of our Artificial Intelligence professor from March 2019 - June 2019. The project has since been accepted into 2019 IEEE Global Conference on Signal and Information Processing (GlobalSIP) and published on IEEE's website.

Some of the technologies/skills used:

• Python

• Matplotlib

• Pandas

• Tensorflow

• Keras

• Jupyter Notebook

Some of my major contributions:

• Created experiments that varied the length, variety, and spacing of manipulations to test model robustness

• Designed functions for applying manipulations to video frames by leveraging various Python image libraries

• Improved the model to a high of 99.6% classification accuracy during testing on limited hardware via use of Residual Networks