Deep Fake Image Detection

Nov 2022

This was a college project to actually use theorethical models in a real life scenario for image classification. Originally I developed my own CNN, defining the layers, features, drop out training strategies etc… I also tried out different training techniques from structuring the training data in randomized batches of different sizes to work better with the optimizer (Adam, or a simple gradient descent), to pre-processing the data via transforming and re-coloring. Different combinations of hyperparameters was tested via GridSearch and get the best performing model.

The scores of my own model were not great, but the experience of developing a pipe-line for machine learning and attempt to design my own model taught me a lot. I also adapated my pipe-line with VGG19 model. I chose this one after a few trials based on my hardware and time limitations. I then tweaked it to fit more on the dataset while making sure not to overfit, to achieve 80% accuracy on the testing data and ranking top 5 out of 20 for my model in the 2 weeks of time constraint I had. (Unfortunately I cannot release the original file since it would violate university rules for giving answers)