MaD-CoRN: an efficient and lightweight deepfake detection approach using convolutional reservoir network

In the modern digital era, human face lies at the core and forms the very basis of any social interaction and communication. It embeds highly precious information alongside facial identities and expressions, but at same time this very trait makes it vulnerable to manipulation attacks. The proliferation of large-scale public image databases alongside advancements in AI-driven image synthesis has heightened the prevalence of “DeepFakes”. While conventional fake detection classifiers suffer from low accuracies due to their susceptibility to manipulation techniques, state-of-the-art convolutional neural network (CNN) models offer high accuracies at the expense of extensive training and computational resources. To address these challenges, we introduce MaD-CoRN i.e. Manipulation detection by Convolutional Reservoir Networks. It is a novel and efficient combinatorial architecture that enhances the feature extraction capabilities using pre-trained convolutional networks with lightweight reservoir computing (RC), an improved form of RNN learning. This approach improves the separation and learning of facial features, resulting in notable speedups and a relative increase of over \(15\%\) $15\%$ in fake face detection accuracy, F1-score values when simulated using “Real and Fake Face Detection (RFFD)” and “100K-Generated fake image” public datasets. MaD-CoRN offers several advantages: (i) it is lightweight and cost-effective, (ii) leverages transfer learning and RC-based feature extraction, (iii) it achieves enhanced accuracies even with small input datasets ( \(\approx \) $\approx$ 1900 samples) and it boasts a highly flexible and generic architecture. Additionally, we propose a unified framework based on various adaptations of the MaD-CoRN architecture.