Deep Fake Images and Videos Detection Using Deep Learning Techniques

Abstract

The rapid evolution of deepfake generation techniques, driven by advanced deep learning models such as Generative Adversarial Networks (GANs), has created significant challenges in ensuring media authenticity and digital trust. Traditional detection approaches, primarily based on convolutional neural networks (CNNs), rely on identifying visual artifacts and inconsistencies; however, their performance degrades when exposed to high-quality or previously unseen deepfake content. Additionally, these methods require large labeled datasets and high computational resources, limiting their scalability and real-world applicability. To overcome these limitations, the proposed mechanism introduces a robust and generalized deepfake detection framework that integrates hybrid deep learning models and lightweight detection strategies. The system combines Capsule Networks with Long Short-Term Memory (LSTM) to capture spatial and temporal inconsistencies, along with NoiseScope-based blind detection to identify intrinsic GAN-generated noise fingerprints. Furthermore, the model supports IoTenabled deployment, allowing efficient execution on resource-constrained edge devices. This multi-level analysis enhances detection accuracy while reducing dependency on extensive training data. Experimental insights indicate that the proposed approach significantly improves generalization, computational efficiency, and robustness against emerging deepfake techniques, making it suitable for real-time and scalable applications in cybersecurity and digital forensics