作者
S. P. Santhoshkumar,G. M. Sathyaseelan,Sai Krishna Gaduputi Subbammagari,Bharath Reddy Bethi,Saurabh Chandra,S. Gajalakshmi
摘要
With the help of sophisticated artificial intelligence (AI), the so-called deepfakes, both photos and videos, are becoming more and more real and pose a threat to the credibility, privacy, and integrity of content of the population. Although very advanced, current deepfake detectors are typically overly biased towards either spatial or temporal delays, are typically based on intensive computing architectures, and lack interpretability, which limits their usefulness in real-world applications, resilience and ability to apply to a wide range of lives. In order to fill the abovementioned gaps, this paper suggests a hybrid neural network to preserve content integrity and to screen real-time deepfakes. The proposed architecture will be a hybrid of Long Short-Term Memory (LSTM) sub-model that should be able to detect temporal distortions, e.g., irregular facial movement and abnormal motion pattern, and a MobileNet backbone that can help extract spatial features in a more efficient way. When trained on a large and heterogeneous dataset of authentic and manipulated samples, MobileNetLSTM framework is based on the state-of-the-art accuracy, which is 97% with its precision, recall, and F1-scores of 93, 94, and 93, respectively. Due to the lightweight design, the proposed model can be deployed on a resource-constrained platform with low-latency inference, which can be used in digital forensics, social media theaters, and content checks. Also, explainable AI (XAI) approaches based on Grad-CAM heatmaps are included to identify anomalous regions on the faces that impact model outputs, which increases clarity and trust in the users. All in all, this work offers a high-performance, scalable, and interpretable way to enhance the authenticity of digital media, and future research should focus on multimodal deepfake analysis, continuous learning to keep up with new forms of manipulation and be more resilient to the new generative models.