Deep Fake Detection
Daksh Baveja
Department Of Computing Technologies
SRM Institute Of Science & Technology, Kattankulathur Chennai, Tamil Nadu db4502@srmist.edu.in
Dr. Nagadevi S.
Department Of Computing Technologies
SRM Institute Of Science & Technology, Kattankulathur Chennai, Tamil Nadu nagadevs@srmist.edu.in
Abstract—The following paper considers an in-depth study of face detection and classification using a pre-trained VGG16 model with a prime focus on separating real from fake facial images. Face detection is a very fundamental task in computer vision and of key importance in various security- and biometric identification-related applications, social media, and so on, in which the above-mentioned Dortania et al. findings will find their use. The idea is to use transfer learning by tuning an already trained VGG16 that was developed for large-scale image classification to do well in a specific task of face authenticity verification.
For this purpose, we constructed a custom dataset with images labeled either ’real’ or ’fake’, sourced from different environments to make it diverse and hence robust. The dataset was then preprocessed by face detection using Haar cascades, resizing, normalization, and augmentation to increase the model’s capacity for generalization. This dataset was trained as well as tested on the modified VGG16 model, where only one fully connected layer at the end was changed to give an output in two classes—one for the real faces and another for the fake ones.
Model performance was ascertained through training loss and accuracy in the training phase. For the 30 epochs of training, the model achieved very good training accuracy. Further performance fluctuation analysis at different epochs used detailed plots of the loss and accuracy. Testing validates further that the model is robust, having a high testing accuracy to ensure the model generalizes on unseen data.
Our results show the effectiveness of transfer learning using VGG16 in face classification, where accuracy was high for the classification of real and fake faces. Thus, this study not only demonstrates the potential of pre-trained deep models in specialized applications but also shows the proper quality of the dataset and its preprocessing towards the attainment of optimal model performance. This trained model is, therefore, deployable in every real-world application where verification of faces is very important, bringing in a reliable tool for improving security and authenticity in digital relations.
Index Terms—deep fake, detection, artificial intelligence, ma- chine learning, digital forensics
