Published Paper Details:

The proliferation of deepfake videos synthesized through deep generative models such as Generative Adversarial Networks (GANs) and neural autoencoders poses critical threats to digital security, media integrity, political stability, and public trust. Existing detection approaches either focus exclusively on spatial artifacts within individual frames using CNNs, or on temporal inconsistencies using recurrent networks, but rarely integrate both. This paper presents a novel hybrid deep learning framework combining Convolutional Neural Network (CNN) and Bidirectional Long Short-Term Memory (BiLSTM) for robust, video-level deepfake detection. The proposed system employs EfficientNetB0 as the spatial feature extractor to capture per- frame facial artifacts, while a two-layer Bidirectional LSTM models temporal inconsistencies across a sequence of 20 uniformly sampled frames. Face regions are detected using the OpenCV Haar Cascade detector with 25% bounding box padding. The model is trained and evaluated on the FaceForensics++ benchmark comprising 7,000 videos across six manipulation categories: Deepfakes, Face2Face, FaceShifter, FaceSwap, NeuralTextures, and DeepFakeDetection. Class imbalance (1:5 real-to-fake ratio) is addressed through Weighted Random Sampling, label smoothing, and stratified splitting. Training employs AdamW optimization, Cosine Annealing scheduling, and PyTorch Automatic Mixed Precision on an NVIDIA Tesla T4 GPU. Experimental evaluation yields a test accuracy of 92.4%, F1-Score of 91.7%, and AUC- ROC of 0.967, demonstrating the superiority of joint spatial- temporal learning over frame-only and temporal-only approaches. A Flask-based real-time web application is developed for practical deployment, providing binary REAL / FAKE verdicts with confidence scores and REST API integration for downstream systems.