Description of Invention
In this research project, python is being used as the main programming language, a preeminent language in machine learning. Speaking of datasets, this is quite important for any Machine learning ng model, hence the datasets that are used in this research are Face Forensics ++ and Celeb V2. Face Forensics ++ is the most popular dataset available with more than 291 papers written and published. This dataset is sourced from 977 YouTube videos. As for Celeb-DF (V2), this dataset contains 590 original videos from YouTube with subjects of different ages, ethics groups and genders and 5639 Deepfake Videos. The proposed method for this research will mainly use the GAN and NAS model. The flow of the proposed method will be firstly, in the data preparation process where the program will copy the video and image files from multiple databases which are the two datasets (Celeb V2 & Forensics ++) and extract the audio files which can be used for multimodal analysis. Next, will be the feature extraction process whereby using opencv to extract frames from videos/images to prepare to convert temporal video data into spatial image data making this suitable for image-based deep learning models and resizing the frames and standardizing to 64 x 64 pixels to ensure efficiency during the training later. After that, it will be data splitting to ensure the model is evaluated on unseen data providing a realistic assessment of its performance and preventing overfitting. Following by the most important part will be where the generator model the GAN Model will generate fake images from random noise which is used to train the discriminator to differentiate between real and fake images. After that, it will combine RGB and Grayscale modalities which is part of the multimodal fusion and later NAS Model is also used to find the best neural network architecture and then save the best model for evaluation and testing later.