Authors: Professor Vikram Singh, Research Scholar Naresh Kumar
Abstract: The rapid generation of deepfake videos poses significant challenges to real-time detection systems, which often encounter a wide range of quality variations and novel manipulation techniques. In order to deal with the challenges faced we propose a hybrid InceptionV3-LSTM network that encapsulates spatial feature extraction and action sequence over time. This architecture employs InceptionV3 pretrained on ImageNet for modeling frame-level artifacts, then stacks a bidirectional LSTM to jointly learn the temporal problem from multiple sequences. The model displays robustness under high compression and when tested on low compression, it shows a decline of just 2.5% in accuracy which implies that the model is sensitive. According to temporal analysis, LSTM proves capable of detecting manipulations as illustrated due to reconstruction loss variance being higher by 3.2× of low-fidelity deepfakes. The deepfake inference data under tool-based evaluation shows this detection. The InceptionV3-LSTM’s full-fledged version achieves 150 ms latency and 96% accuracy while real-time performance granted 20 ms/frame latency using MobileNet-LSTM with no accuracy loss (96%). The solution presented in this study can be deployed, balancing detection accuracy and computational efficiency. However, limitations still exist for novel manipulation techniques and multi-modal integration could be an avenue for future work. The objective of this experimental study is to achieve the enhanced detection accuracy using separate training and testing datasets of deepfake videos which can help get a better security performance and robustness. By integrating these advances, the application allows users to check the authenticity of people and deep fake videos. A framework for real-time detection of visual deepfake videos analyzes pixel-level artifacts including blending boundaries, double edges, eyes, ears, nose, head pose inconsistencies and other possible parameters sub-pixel noise patterns unique to manipulation videos.
