Deep Learning Methods for Real -Time Detection of Deepfake Faces in Images and Videos
Jyothi Chinna Babu, Alind, Mallikharjuna Rao Nuka, C. Venkatesh, Tanmay Kumar, Adarsh Kumar · 2025
This paper presents a Deep Learning Methods for Real- Time Detection of Deepfake Faces in Images and Videos which employs a hybrid deep learning model, consisting of a Long-Short Term Memory (LSTM), ResNet50, and Convolution Neural Networks (CNNs). A CNN architecture was built to extract strong spatial information from facial images, allowing for infections in the identification of extremely subtle abnormal features such as artificial lighting, irregular facial textures, and pixel-level anomalies found in deepfake media. Further, it utilizes a deep residual network, ResNet50. Such a deep network allows the machine to embed higher-level features and render subtle manipulation artifacts viable for detection, possible to overlook in other elementary CNN models. LSTMs are used in the present work to capture long-term temporal dependencies among various identified frames. This Model receives training from many datasets containing collections of diverse real and synthetic facial images and video. It also uses preprocessing techniques to make certain that the model generalizes well in various situations of lighting, poses, or video compressions. This includes face alignment, noise reduction, data augmentations-an increase that directs filtering. Over 93% accuracy was achieved on all benchmark datasets for real-world deepfake detection tasks, using a hybrid CNN-ResNet50-LSTM model. This work shows the significant benefit coming from the integration of temporal and spatial analysis that can practically be applied for security, ethical media authentication, and content verification.