Deep Learning for Facial Forgery Detection Performance Evaluation of DenseNet201, InceptionV3 and ConvNeXt

The recent spread of AI-generated face forgery is one of the greatest threats to visual media credibility. The Proposed work compares three deep transfer learning models DenseNet201, InceptionV3, and ConvNeXt, in detecting manipulated facial images. An 8,000 real and fake facial image dataset was used to train and benchmark models under consistent experimental condition. ConvNeXt achieved the best classification accuracy of 91.25 % which is much higher than that of DenseNet201 (75.12 %) and InceptionV3 (68.38 %). In addition, ConvNeXt had better trade-off between true positive and false positive rates, which means better generalization and resistance to overfitting. These findings prove the applicability of ConvNeXt in the robust detection of facial forgery and highlight potential application in the practical implementation of the facial authenticity determination. Future research will investigate ensemble methods and real-time inference.