Deep Fake Detector
CNN Detector
Our research is aimed at exploring whether a trained model detector is able to differentiate real photos from fake photos. We thus ask the question: What is the probability that a fake image detector will be deceived by fake images? And, by analyzing how the trained model identifies real and fake images, we also want to achieve a more practical contribution: how can the everyday Internet user lookout for false images (for instance, what features commonly give the falsity away)? As we answer these questions, we will also explore the ethical issues related to the use of fake imagery, especially with regards to issues of identity and trust.
Client
Services
ML AI
Industries
AI
Date
April 2021
The dataset used for the initial training and testing of the model is from a Kaggle Deepfake Detection Challenge. To evaluate the performance of the model, two key metrics are used: Precision and Recall. In the convolutional neural network, there were 2 variables that had to be selected and varied in an attempt to optimize the results: image size and epochs/iterations. The images loaded in the model were square in shape, and 4 different sizes for the length of the square were chosen: 32, 64, 128, and 256 pixels.
