A Hybrid Approach for Video Forgery Detection in Digital Media Using Inception v3

Numerous video forgeries based on AI-driven media manipulation techniques continue to increase in popularity which creates substantial risks for digital forensics as well as security systems and validates media authenticity. The detection methods for forgeries using both statistical analysis and manual features can no longer identify leading-edge AI-generated modifications thus requiring improved automatic detection systems. Deep learning-based models present themselves as an effective solution to handle this problem. Conferral Inception v3 along with other CNN systems shows excellent capabilities in extracting spatial information from singular video frames yet struggles to evaluate temporal inconsistencies which represents an essential factor for detecting modified sequences. The proposed study presents a combined video forgery detection system that uses Inception v3 spatial features combined with LSTM along with SVM and anomaly detection algorithms for temporal video examination. The proposed method surpasses standard CNN detection methods because it uses temporal dependencies between video frames for better accuracy. The hybrid model shows superior outcomes than standalone CNN models ResNet and VGG regarding deepfake manipulation and frame tampering detection because it achieves better precision, recall and AUC-ROC scores. The use of Grad-CAM heatmaps together with confusion matrices allows to obtain detailed information about forgery patterns. Real-time deployment of the model becomes possible through optimization techniques that include quantization and pruning while the system demonstrates its functionality for digital forensics work as well as social media moderation and misinformation detection. The proposed system framework advances the creation of efficient scalable and resistant video forgery detection systems to safeguard media security.