Deepfake Detection with Metaheuristic Algorithms and Deep Features

Abstract

The rapid development and spread of deepfake technology have posed serious threats to cybersecurity, information security, privacy, and public safety; consequently, reliable detection of deepfake content has become a critical necessity.   We present a metaheuristic-based hybrid approach that combines deep learning architectures with metaheuristic algorithms. In the study conducted on FaceForensics (FF++) and Celeb-DF datasets, four feature vectors extracted from Xception and ResNet50 architectures underwent a metaheuristic-based feature selection process incorporating Cuckoo Search Algorithm (CS), Success-History Based Adaptive Differential Evolution with Linear Population Size Reduction (L-SHADE), Particle Swarm Optimization (PSO) and Weighted Mean of Vectors Optimization Algorithm (INFO) algorithms. Sub-feature vectors were obtained through feature selection for each algorithm. The four feature vectors obtained from the architectures and sixteen sub-feature vectors generated after feature selection were classified using machine learning and deep learning methods. When comparing performance metrics before and after feature selection, the INFO algorithm provided the highest performance across both datasets, achieving AUC values of 99.05% for the FF++ dataset and 99.01% for the Celeb-DF dataset. We believe that our comprehensive experimental study demonstrates more significant and effective results than existing methods.