Deepfake Image Detection Using ResNet50 Model

Deepfake images, created through advanced AI techniques, pose significant cybersecurity risks, facilitating identity fraud and potentially damaging reputations and financial stability. The primary challenge in detecting deepfakes lies in ensuring algorithms generalize well across diverse datasets, minimizing errors like False Positives (FP) and False Negatives (FN) on unseen data. This study focuses on enhancing deepfake detection using the ResNet50 model and Gaussian blur preprocessing to bolster model generalizability. The methodology encompasses five phases: Data Collection, Data Splitting, Data Pre-processing, Training, and Model Testing on the DeepFakeFace and Deepfake Processed datasets. Evaluations, including accuracy, specificity, recall, and precision, are conducted on 4000 images per dataset split in ratios of 60:20:20 and 80:10:10. Results indicate superior performance with the 80:10:10 split, achieving peak accuracy of 82.75%. Gaussian blur notably enhances specificity, increasing from 54.67% to 65.00% with the 80:10:10 ratio, underscoring its role in improving model robustness across varied datasets. This research highlights critical advancements in cybersecurity measures against deepfake threats.