DEEP VISION IN DETECTING VISUAL DECEPTION

The intensive development of deep learning made it possible to produce highly realistic manipulated content, such as face-swapped videos, forged images, and synthetic audio, which are very dangerous in the form of misinformation, identity fraud, and other criminal activities. This paper introduces Deep Vision AI, a multi-modal deep learning model of identifying manipulated media with the main emphasis on the production of fake content. The suggested system combines video, image, and audio analysis with such sophisticated models as Xception with Long Short-Term Memory (LSTM) to model the video sequence, EfficientNetB0 to analyze images in the context of video forensics, and MFCC-based to extract features and classify audio with the help of the ASVspoof dataset. Several benchmark datasets, such as FaceForensics++, Celeb-DF and DeepFake Detection datasets are merged to improve generalization and strength. The results of the experiment show that the proposed system has an accuracy of 92, 86 and 87 percent on video, image and audio respectively and has a total system accuracy of 90 percent with a majority voting fusion mechanism. The system is deployed as a web-based application through the use of Flask allowing real-time identification of manipulated media. The findings are that multi-modes integration greatly enhances reliability of detection relative to single-modes.