A Comparative Study of DenseNet121 and InceptionResNetV2 model for DeepFake Image Detection

The rise of deep fake technology poses a significant challenge to societal integrity, necessitating robust detection methods. This research study investigates the effectiveness of DenseNet121 and InceptionResNetV2 deep learning models in detecting deep fake images. The study starts with the curation of the dataset and then applies different preprocessing methods to improve the quality of the dataset like scaling, normalization, and data augmentation. The DenseNet121 and InceptionResNetV2 models are trained to differentiate authentic and manipulated images based on distinctive features. DenseNet121’s dense connectivity facilitates feature reuse, enabling effective extraction of discriminative features crucial for identifying manipulated content. In contrast, InceptionResNetV2’s sophisticated architecture exploits both inception and residual connections, enhancing its ability to capture intricate patterns indicative of deep fake manipulation. The comparative analysis of performance metrics reveals InceptionResNetV2’s superiority, achieving an impressive accuracy rate of 99.78% compared to DenseNet121. These findings underscore the efficacy of leveraging existing architectures for identifying deep fake images, thereby contributing to the ongoing efforts to combat misinformation and safeguard digital content integrity.