Hyperparameter Optimization for Deepfake Detection Improving CNN and Transformer Performance Using Bayesian Search

We provide an extensive study of hyperparameter optimization for CNN and Transformer models using the deepfake detection problem. We will use Optuna for Hyperparameter Tuning with Bayesian optimization in which we will tune some important parameters (dropout rate, learning rate, optimizer type). Experiments are performed with a real and fake image dataset extracted from video clips. We evaluate the optimized models in terms of accuracy, precision, recall, F1-score, and ROC-AUC metrics. With hyperparameter discovered by bayesian search, the Transformer model presta the best performance where it attained 94% accuracy and 0.94 AUC compare to optimized CNN for which we get 90% accuracy, 0.89 AUC. Moreover, we observe that Transformer has a better recall to identify fake videos (96% vs 83% for CNN), indicating its robustness in identifying deepfakes. In conclusion, the proposed framework for hyperparameter optimization can be a valuable addition to the growing body of research aimed at developing robust and effective models for deepfake detection.