EXPENDABLE DEEPFAKE DETECTION USING MULTI-STAGE DEEP LEARNING WITH SPATIAL, TEMPORAL, AND FREQUENCY FORENSIC PIPELINES

Deepfake technology uses AI to create realistic but fake images, videos, and audio based on existing media. While intriguing, it poses significant threats in the digital era, affecting reputations, spreading rumors, and influencing political opinions. Advances in deepfake generation make it more convincing and accessible, increasing its misuse in cybercrimes such as identity theft, cyber extortion, fake news, financial fraud, and blackmail. To combat these threats, social media and networks seek intelligent algorithms for deepfake detection. The sophistication of deepfake technology is constantly increasing; therefore, robust and explainable deepfake detection is indispensable for digital forensics. Most existing approaches to deepfake detection focus on single-domain features, such as spatial inconsistencies, and have poor generalizability over different datasets. Moreover, they seldom handle artifacts produced by generative models like StyleGAN3, such as fine-grained blending errors and GAN-induced high-frequency noise. To overcome these limitations, we introduce an Explainable Deepfake Detection Framework that integrates a multi-stage deep learning pipeline with spatial, temporal, and frequency feature extraction. Our model begins with data collection using FaceForensics++ combined with synthetic deep fakes generated via StyleGAN3, guaranteeing diversity across compression levels, ethnicities, and poses while mitigating bias. Pre-processing employs MTCNN for face alignment and DWT for frequency domain analysis, enhancing sensitivity to subtle artifacts. Feature extraction uses three specialized modules: (1) Xception CNN for spatial features to detect blending artifacts and edge inconsistencies, (2) LSTM-based Temporal Network to capture unnatural motion artifacts over video frames, and (3) DCT-DenseNet to identify high-frequency inconsistencies in frequency space. The multi-stage fusion classifier combines the features using an Attention-Based Weighted Fusion strategy to optimize accuracy through an emphasis on influential modalities. Grad-CAM and SHAP post-processing will provide explainability by showing regions that contribute to the artifacts and quantifying the importance of features. Experiments on the FaceForensics++ dataset achieved 99.2% accuracy, 98.7% F1 score, and 0.995 AUC-ROC, making it state-of-the-art. This work not only enhances the accuracy of detection but also improves interpretability, allowing forensic experts to understand and trust deepfake predictions better in the process.