Refining digital security with EfficientNetV2-B2 deepfake detection techniques
G. N. Vivekananda, T R Mahesh, Muskan Gupta, Arastu Thakur, Anu Sayal · Egyptian Informatics Journal · 2025
The rise in digitally altered images has made research on robust solutions for real image verification across sectors, including media and cybersecurity very essential. Deepfake technology’s development compromises digital media’s validity and calls for advanced detection to address. With EfficientNetV2-B2, a novel improvement in convolutional neural networks that is considered efficient and effective, the present research proposes a strong method for separating deepfake and real images. To ensure equal ratio, the paper utilized a balanced dataset consisting of 100,000 photos divided equally between real-world and deepfake classes. Methodology involved image preprocessing to the same dimensions, model strength augmentation techniques, and a rigorous training process with parameter optimization for precision. Interestingly, the study employed an independent learning rate adjustment method for enhancing training performance, resulting in better model calibration. Experiment setup results showed a staggering 99.885 % in classification accuracy and a corresponding high F1 score, thereby establishing the capability of the model in deepfake detection. Extensive exploration also confirmed there were evident cases of misclassification, which indicated areas where training model and image processing procedures should be improved. The results illustrate the prospect of applying EfficientNetV2-B2 in situations where high accuracy is needed in photo verification.