MDR-LOD2 Model: Forgery Detection using Modified Depth ResNet features and Layer Optimized Dunnock Deep Model from Videos

Abstract

Digital forgery detection implies the identification of any modifications or manipulation of the digital content, typically image, video, or document, to confirm their authenticity. Consequently, this contribution seeks to address the challenges experienced by existing techniques by introducing the Modified DepthResNet descriptor and Layer Optimized Dunnock Deep model (MDR-LOD2) model. MDR descriptor is proficient at generating features in ResNet architecture and hence it helps in the fusion of a DepthNet to detect depth-related cues which plays a crucial role in spotting forgery. More specifically, the MDR descriptor captures the subtle details via the spatial connections and depth perception, resulting in boosting the detection performance. The hybrid optimizer strategy combines meticulous exploration and dynamic adaptation together increasing the model's ability to detect splicing forgery. The proposed approach exploits the LOD2 architecture well suited for capturing the temporal aspects and effectively analyzes the intricate patterns of video data. Additionally, the LOD2 model is enabled with the Dunnock Hunt Optimization (DHO) algorithm for layer optimization facilitating optimal performance of every layer in LSTM. Moreover, the integration of LOD2 and MDR descriptor in conjunction with the DHO algorithm in the proposed approach assist in identifying the forged regions in the video frames. The experimental results demonstrate that the proposed approach attains an accuracy of 98.54 %, sensitivity of 98.54 %, specificity of 98.53 %, and F1-score of 98.54 % for DSO-1. For DSI-1 DTS, the proposed approach achieves remarkable results with high accuracy of 98.47 %, sensitivity of 98.41 %, specificity of 98.52 %, and F1-score of 98.47 %. Finally, the proposed model obtained the remarkable results for the Face Forensics database achieving high accuracy of 97.83 %, sensitivity of 97.76 %, specificity of 97.89 %, and F1-score of 97.83 % outperforming other existing techniques.