About me
Omar's Bio: I am a Senior at the University of Southern Mississippi majoring in Computer Science. I conducted research on using deep learning algorithms to combat the spread of AI-driven misinformation, using open-source datasets to test widely used deep learning models. I plan to continue testing other algorithms on datasets more specific to deepfakes. In my free time, I enjoy playing video games, watching TV shows, riding my bike, and other activities.
Omar's Research Abstract: The rise of Generative AI allows malicious actors to create highly convincing false images (deepfakes) at minimal cost, raising concerns about the spread of misinformation. Various machine learning classifiers and deep learning models have been utilized to combat this issue. This study compares probabilistic classifiers, Naïve Bayes (NB) and cross-validation, with a Convolutional Neural Network (CNN), a deep-learning model, on two datasets: MNIST (60,000 digit images) and EMNIST (131,600 alphanumeric images). NB was evaluated as is and with 10-fold cross-validation, while the CNN used transfer learning and ten training epochs. On MNIST, NB achieved an accuracy of 81.1% and an F1 score of 81.51%, confirming its suitability as a lightweight baseline. With 10-fold cross-validation, accuracy slightly improved to 84.25%/84.39%. The CNN outperformed both machine learning methods with a 92.80% evaluation accuracy on the EMNIST dataset, along with a precision of 92.7%, a recall of 92.8%, and an F1 score of 92.69%. These results show that deep learning models remain state-of-the-art for visual misinformation tasks and are often the basis for most newer models, which are either CNN-based or hybrid. Future work will test the comparison on domain-specific deepfake datasets.
