Stacked multi-fusion CNN: an adaptive attention model for privacy preserving deepfake forensics

The emergence of Generative Artificial Intelligence (Gen-AI) and Generative Adversarial Network (GAN)-based deepfakes poses significant security risks in sociocultural and sociopolitical domains. This necessitates the development of advanced and effective detection methods to prevent vulnerability in social networks. Classical Machine Learning (ML) algorithms have their limitations, especially in classifying the deepfakes. To address these issues, this paper suggests a privacy-preserving Stacked Multi-Fusion (SMF) Convolutional Neural Network (CNN) approach to classify deepfakes. An improved CNN model is proposed, integrating an adaptive multi-scale attention framework with enhanced residual blocks and a Squeeze-and-Excitation (SE) mechanism. The selection of these components is backed with an ablation study to report the individual contribution to the overall optimal architecture. A hybrid lossless multilayer cryptosystem based on a chaos-based approach, Deoxyribonucleic Acid (DNA)-based computing, etc., is developed to secure images in cloud storage. The efficacy of the proposed SMF model is validated using the 140K Real and Fake Faces (RFF) image dataset. The proposed approach was found to achieve stable performance with minimal variation in various tests. It achieved a test accuracy and ROC-AUC of 97.80 and 99.79, respectively. This paper provides comprehensive relevant factors for building effective synthetic media detection systems.