Attention-Enhanced CNN for High-Performance Deepfake Detection: A Multi-Dataset Study

Deepfakes, which emerge from advanced deep learning techniques, present complex ethical and security challenges across media and communication landscapes. While offering creative potential in education and entertainment, synthetic media technologies simultaneously threaten societal trust through potential misinformation, opinion manipulation, privacy violations, and identity fraud. With the advancement of deep learning models, the creation of deepfake images has become easier and more convincing, resulting in the development of reliable deepfake detection models. This research works with a method that combines multi-head self-attention (MHSA) with a custom designed convolutional neural network (CNN) to develop a robust deepfake detection model. We created a dataset called the Center for Cyber Defense DeepFake (CCDDF) dataset by generating fake images using publicly available Artificial Intelligent (AI) tools and trained our model on these data, achieving a detection accuracy of 97% and an AUC score of 99.58. Additionally, we evaluated our model on the 140K Real and Fake Faces dataset and the Celeb-DF v2 dataset, where it demonstrated exceptional performance with accuracies of 98% and 94% respectively, and corresponding AUC scores of 99.75 and 98.72. To enhance interpretability, we utilized attention heatmap visualizations to analyze the decision-making process of the model. Our results demonstrate the effectiveness of combining multi-head self-attention with a convolutional neural network for deepfake detection, highlighting its strong performance across multiple datasets and its potential for real-world applications.