International Journal of Science, Engineering and TechnologyAuthors: Indore Tanuja Sunil, Bhosale Sachin Bajirao, Bangar Abhishek Popat
Abstract: The rapid advancement of artificial intelligence and deep learning has made it in-creasingly easy to create realistic manipulated images and videos, commonly known as deepfakes. Such manipulated content can create serious risks to digital security, privacy, identity protection, and the reliability of online information. Existing deep-fake detection methods mainly focus on spatial features, frequency artifacts, or temporal inconsistencies, but their performance may decrease when dealing with compressed videos, unseen manipulation techniques, and cross-dataset conditions. This paper proposes a Hybrid Spatial-Frequency-Temporal Siamese Deep Learning Framework for robust deepfake detection. The proposed approach combines spatial features extracted from RGB facial frames using a convolutional neural network, frequency-based features obtained using Error Level Analysis (ELA) and frequency analysis, and temporal features learned from sequential video frames using a bidirec-tional recurrent model. A Siamese learning mechanism is incorporated to improve the separation between authentic and manipulated representations, inspired by recent forgery verification research. The extracted features are fused and classified into Real and DeepFake categories. Grad-CAM is also considered to provide visual expla-nations of the model's decisions. The proposed framework aims to improve detection robustness, generalization, and interpretability under different manipulation and compression conditions.
