Performance evaluation of deep learning models for detecting deep fakes


  • Aishwarya Rajeev
  • Raviraj P



Face Forensics, Convolutional neural network, recurrent neural network, DAFDN, Resnet v2, Xceptio


The proliferation of deep fake content in multimedia has necessitated the development of robust detection mechanisms. In this study, a comparative analysis of four state-of-the-art deep learning models for detecting deep fakes is conducted: CNN+RNN, DAFDN, Hybrid Inception ResNet v2, and Xception. The evaluation focuses on their performance metrics, emphasizing accuracy as a primary measure. Through extensive experimentation and evaluation on a comprehensive dataset, the findings reveal notable distinctions among these models. The CNN+RNN architecture demonstrates a commendable accuracy of 94.8%, providing a solid baseline for comparison. Surpassing this, the DAFDN model achieves an accuracy of 97.8%, showcasing superior discriminatory capabilities in identifying manipulated content. Furthermore, the CNN model stands out with an accuracy of 98%, exhibiting remarkable effectiveness in distinguishing between genuine and deep fake media. The comparative analysis delves into the strengths and weaknesses of each model, shedding light on their respective performance levels in detecting sophisticated deep fake content. The observed accuracies underscore the nuanced differences in their architectures and training methodologies, offering insights crucial for selecting appropriate models based on specific detection requirements.