Enhancing deepfake detection with dual-strategy Puma optimization and stacked long short-term memory models

Abstract

A deepfake is a synthetic media model used to modify, manipulate, and generate videos, images, and audio recordings. However, the existing fake detection algorithm struggles to differentiate real from fake content due to redundant information in the extracted features, which degrades its performance. To overcome this limitation, a dual-strategy-based Puma optimization(DSPO) algorithm feature selection model is proposed to identify significant features that effectively distinguish normal from fake images. These features are then processed by a stacked long short-term memory (SLSTM)-based detection model to differentiate between real and fake videos. Initially, videos are acquired and preprocessed into image frames, and relevant features are extracted using transfer learning-based approaches. Next, the proposed DSPO algorithm employs a dual strategy, involving a heavy-tailed distribution and adaptive parameter tuning, to address the algorithm’s limitations and select an optimal feature set. Finally, the SLSTM model uses a sech activation function to identify fake videos and improve deepfake detection accuracy. Experimental results demonstrate that the DSPO-SLSTM model achieved accuracies of 99.54%, 99.75%, and 99.54% on the Forensic Face (FF++), Celeb-DF, and Celeb-DF-V2 datasets, respectively, which is greater than existing detection models such as discrete wavelet transform-vision transformer (DWT-ViT).