Enhancing Fake Face Detection with Meta-Learning and Ensemble Methods

Abstract

"Nowadays, facial recognition technology has improved significantly, greatly enhancing security, personalized customer experiences, and reliable biometric verification. But the emergence of deepfake technology, where AI can generate images that look very real, is a concerning issue. Deepfakes have been detected primarily through slow and unreliable visual clues. Currently, image processing and deep learning techniques are being used to identify fake faces quickly and accurately. Identifying deepfakes early is imperative to prevent their abuse and enhance security. IoT devices can rely on deep learning algorithms, which offer real-time data analysis and precise fake face detection. These algorithms use large amounts of data and sophisticated neural networks, becoming extraordinarily accurate at distinguishing real from fake images. This paper proposes a model that uses a meta-learning and ensemble approach for deepfake face detection. Atfirst, we have implemented five different deep learning models: EfficientNetV4, MobileNetV2, ResNet50, NASNetMobile, and DenseNet. These models provide detection accuracies of 89%, 76%, 79%, 83%, and 84%, respectively. Subsequently, we applied stacking and blending with meta-learning, using a random forest classifier as the meta-learner. Stacking achieved 87.40% accuracy, while blending reached 92.46%, demonstrating the effectiveness of our approach for fake face detection. The high accuracy of the proposed system enhances security and ensures applicability on various devices, thereby enabling large-scale security deployment."