An Real-Time Deep Fakes and Face Forgery Using Transfer Learning Algorithm

A novel deep learning architecture for deepfake detection was proposed that combines LSTM and CNN methods, implemented using Python on the Kaggle platform. The model was evaluated using two datasets: DFDC and Ciplab. Dataset preprocessing involved 19,148 real images and an equal number of fake images, with 80% allocated for training using 128 × 128 image sizes. Binary cross-entropy function 5.4 was used to calculate error rates during training iterations. The results demonstrated high accuracy rates of 97.32% and 98.24%, with low error rates of 0.15% and 0.26% for the respective datasets. These results confirm the model's adaptability for accurate deepfake prediction. The effectiveness stems from the complementary strengths of both approaches: LSTMs effectively capture sequential temporal dependencies between frames to identify deepfake inconsistencies, while CNNs evaluate frame content to detect visual artifacts, inconsistencies, or unusual patterns indicating manipulation. The hybrid approach leverages the strengths of both methods, resulting in a more efficient and effective model capable of adapting to various deepfake generation techniques. By combining temporal analysis capabilities of LSTMs with the spatial feature extraction abilities of CNNs, the architecture achieves a more comprehensive deepfake detection system