DeepFake Detection: Evaluating the Performance of EfficientNetV2-B2 on Real vs. Fake Image Classification

Abstract

The surge in digitally altered images has necessitated advanced solutions for reliable image verification, impacting sectors from media to cybersecurity. This work provides an effective method of real vs. deepfake image distinction through utilization of the EfficientNetV2-B2 model, the latest in convolutional neural networks known for its accuracy and effectiveness. The research utilized a big dataset of 100,000 images equally divided between deepfake and real classes to create a balanced sample. The methodology involved preprocessing images to a fixed size, utilizing augmentation techniques to enhance model robustness, and employing a systematic training schedule along with accuracy parameter optimization. Significantly, the research utilized an automated learning rate adjustment mechanism to optimize training performance, contributing to a complex model calibration. Outcome of the experiment design was showing 99.89% classification accuracy and an equally impressive F1 score, which is a measure of the efficiency of the model in identifying deepfakes. The results provided in-depth analysis with some misclassifications, providing recommendations for potential image processing and model training improvements. The outcome points to the suitability of applying EfficientNetV2-B2 where there is a requirement for high accuracy in image authentication.