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

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• Generative Artificial Intelligence (Gen-AI) and Generative Adversarial Network (GAN)-based deepfakes present significant security challenges.
• Classical Machine Learning (ML) methods struggle with accurate deepfake classification.
• Vulnerability in social networks necessitates advanced deepfake detection.

Purpose of the Study:

• To propose a privacy-preserving Stacked Multi-Fusion (SMF) Convolutional Neural Network (CNN) for effective deepfake classification.
• To develop an improved CNN model integrating adaptive multi-scale attention, enhanced residual blocks, and Squeeze-and-Excitation (SE) mechanisms.
• To create a hybrid lossless multilayer cryptosystem for securing images in cloud storage.

Main Methods:

• An improved CNN architecture incorporating an adaptive multi-scale attention framework, enhanced residual blocks, and SE mechanism.
• Ablation study to validate the contribution of each architectural component.
• Development of a hybrid lossless multilayer cryptosystem using chaos-based and Deoxyribonucleic Acid (DNA)-based computing.
• Validation using the 140K Real and Fake Faces (RFF) image dataset.

Main Results:

• The proposed SMF model achieved a test accuracy of 97.80% and a ROC-AUC of 99.79%.
• The model demonstrated stable performance with minimal variation across tests.
• The integrated cryptosystem effectively secured images in cloud storage.

Conclusions:

• The developed SMF-CNN approach offers a robust and accurate solution for deepfake detection.
• The findings contribute to building more effective synthetic media detection systems.
• The privacy-preserving nature and high accuracy make the model suitable for social network security.