Illumination Effects on Deep Fake Detection: Performance Comparison of CNN Architectures

While AI has many advantages, it also has a few drawbacks associated with “deep fakes.” Deep fakes involve the creation of very realistic videos and images realized through technologies that work with facial and voice modalities for deceptive purposes. Researchers have developed various techniques to detect these manipulated videos and images, employing complex neural network models such as convolutional neural networks (CNNs). However, deep fakes are very difficult to identify due to lighting and shadow effects that may hide some signs of manipulation. No one technique of detection could have the ability of identifying the tampered images under all kinds of illumination conditions. This paper compares different CNN models-ResNet50, EfficientNet, and MobileNetV2-to improve detection accuracy so as to be more resistant to new deep fake techniques, in particular under different illumination conditions. The importance of the study lies in arriving at which of the considered CNN models turns out to be most accurate under a specific illumination effect. Such an understanding could, therefore, improve the results of detection even when applied to cases of tampered faces in fake Aadhaar detection and fake id detection, thereby clearly bringing out the usefulness of the research in practice.