Video deepfake detection using Particle Swarm Optimization improved deep neural networks

Video deepfake detection using Particle Swarm Optimization improved deep neural networks

22 February 2024 | Leandro Cunha, Li Zhang, Bilal Sowan, Chee Peng Lim, Yinghui Kong
This article presents a novel approach to video deepfake detection using Particle Swarm Optimization (PSO) for hyperparameter selection in deep neural networks. The authors propose a hybrid EfficientNet-Gated Recurrent Unit (GRU) network and EfficientNet-B0-based transfer learning for video forgery classification. The PSO algorithm is modified to incorporate composite leaders and reinforcement learning-based search strategy allocation to mitigate premature convergence. The system comprises three key steps: data preprocessing for facial region extraction, PSO-based hyperparameter optimization during network training, and model establishment using the selected optimal settings for fake/real video classification. The proposed PSO-based EfficientNet-B0 and EfficientNet-GRU networks outperform existing state-of-the-art methods on several deepfake datasets. The research introduces a new PSO variant that combines adaptive nonlinear functions for composite leader generation and Q-learning for optimal search operation dispatch. The proposed method demonstrates superior performance in identifying deepfakes by effectively extracting spatial-temporal cues and capturing inter/intra-frame inconsistencies. The study also evaluates the effectiveness of the proposed PSO algorithm in solving a variety of benchmark functions, showing statistical superiority over other search methods. The results indicate that the proposed system achieves higher accuracy in detecting deepfakes compared to traditional methods, highlighting the potential of PSO in enhancing deepfake detection performance.This article presents a novel approach to video deepfake detection using Particle Swarm Optimization (PSO) for hyperparameter selection in deep neural networks. The authors propose a hybrid EfficientNet-Gated Recurrent Unit (GRU) network and EfficientNet-B0-based transfer learning for video forgery classification. The PSO algorithm is modified to incorporate composite leaders and reinforcement learning-based search strategy allocation to mitigate premature convergence. The system comprises three key steps: data preprocessing for facial region extraction, PSO-based hyperparameter optimization during network training, and model establishment using the selected optimal settings for fake/real video classification. The proposed PSO-based EfficientNet-B0 and EfficientNet-GRU networks outperform existing state-of-the-art methods on several deepfake datasets. The research introduces a new PSO variant that combines adaptive nonlinear functions for composite leader generation and Q-learning for optimal search operation dispatch. The proposed method demonstrates superior performance in identifying deepfakes by effectively extracting spatial-temporal cues and capturing inter/intra-frame inconsistencies. The study also evaluates the effectiveness of the proposed PSO algorithm in solving a variety of benchmark functions, showing statistical superiority over other search methods. The results indicate that the proposed system achieves higher accuracy in detecting deepfakes compared to traditional methods, highlighting the potential of PSO in enhancing deepfake detection performance.
Reach us at info@study.space