A Novel Dense-Swish-CNN with Bi-LSTM Framework for Image Deepfake Detection

Artificial intelligence (AI) developments have revolutionized technologies and methodologies, particularly for malicious uses, especially since the advent of generative adversarial networks (GANs) in 2014. GANs can fabricate media streams, seamlessly blending them into different environments. Advanced machine learning techniques can also generate highly realistic deepfake images, which can be exploited for harassment and blackmail. Furthermore, deepfakes can manipulate public sentiment and opinions about individuals or groups. This situation necessitates the development of automated methods to detect deepfakes. In this work, we propose a hybrid deep learning approach for deepfake face detection. The base models include CNN+Bi-LSTM, ResNet34+EfficientNet+Bi-LSTM, and MobileNet+EfficientNet+Bi-LSTM. Initially, images from multiple sources were collected and also generated using various GANs, and the proposed Dense-Swish-CNN with Bi-LSTM was trained. Extract spatial & temporal information from input samples using hybrid pre-trained models. We evaluated our model on challenging datasets such as DFDC, CelebDF, ForgeryNIR, and UADFV. Our experimentation demonstrated the effectiveness of our proposed approach with accuracy 97.76%. Additionally, we assessed the effectiveness of deepfake detection method and validated the explainability of the Dense-Swish-CNN model by performing cross-dataset validation.