Content area
Video forgery (VF) is manipulating fake videos by modifying, coordinating or generating new content in the video sequence. The generated fake videos pose a great threat to social stability, enhancing the necessity to introduce fast, efficient video forgery identification techniques. Many existing studies reported effective techniques to detect forged videos at minimal complexity. However, existing techniques failed to obtain global assessments of the entire video frames and are less robust against fast-moving objects. Thus, this article proposes a novel optimized trident encoder-decoder network with adaptive deep learning models for detecting and localizing video forgeries. Initially, the input videos are converted into multiple frames and forwarded to the encoder part of the detection network. In this encoder part, an Attention SqueezeNet (AttSNet) is proposed for obtaining three branches of frames: pristine, forgery and both (pristine and forgery). Then, the bi-directional long short-term memory (Bi-LSTM) model is introduced in the decoder part to accurately detect the presence/absence of forgery from the given video frames. After detection, the forged region is analyzed by proposing a novel Adaptive ResNet (A-ResNet) with a generative adversarial network (GAN) model in the localization process. Here, the features from forgery-detected video frames are extracted by employing A-ResNet and the forged regions are localized effectively by the GAN method. In addition, this study proposes a hybridized wild-hunt (WiH) optimizer technique to update the weight parameters of the proposed model. The proposed method is implemented in the Python platform, and the whole experiment is processed with a face forensic database. In the experimental section, the accuracy of 95.4% and 96.1% and the time complexity of 1.43 s are obtained for detection and localization, respectively.