Problem Identification:

The rapid proliferation of deepfake prevalence proceeds at an alarming rate. Especially now, during the digital age, the need for a deepfake detection system is increasingly evident. Deepfakes are used to disseminate misinformation through harassment, extortion and other malicious activities. In North America alone, Deepfake-related fraud grew by 1,740 percent (2022)—a figure that continues to rise each year. In order to resolve this issue, the progression of a binary deepfake classification framework is necessary. The framework applies a convolutional neural network (CNN) structure on 140,000 training images. The first step is image normalization, where input images are resized to 128x128 pixels and converted into floating point decimals. In a CNN architecture, there are 2 principal categories: pretrained bases such as VGG16 and custom bases calibrated using hyperparameter tuning, and backpropagation. This led to the research question, “How does the use of custom versus pretrained base models impact the binary accuracy of a deepfake binary classification system?” and the hypothesis, “When a custom base is utilized then the binary accuracy of the finished model will exceed that of a model with a pretrained base.” By iteratively testing models with a custom and pretrained base, the ideal CNN model was derived. This model utilizes a custom base for feature extraction, and a custom head for label prediction. Deploying the model on unexposed data, it outperformed the pretrained base model—attaining an accuracy of 95.26%. This innovative solution will combat deepfake fraud, safeguarding individuals’ identities.