Advanced Deepfake Detection Using Honey Badger Optimization and ELM Classifier

Abstract

Deepfake poses a significant threat in contemporary times as it has negative impact on society. It is incredibly difficult to distinguish manipulated faces from real ones, even when scrutinized closely. Deepfake often struggles to replicate natural human expressions and subtle facial movements accurately. Detection methods focusing on inconsistencies in facial expressions, unnatural movements, or mismatches between facial movements and the emotional context can identify manipulated content. The current approaches face challenges in handling post-processing effects such as compression, noise, and changes in lighting. There is a lack of extensive research addressing the detection of both audio and visual deepfake content. This paper introduces a novel model designed to identify deepfake content within video frames. Our model detects deepfake by splitting the videos into frames and extract the features. The spatio temporal features of the frames help us to identify multiple frames. To reduce resource utilization, the fused frames are given as input to the trained model. An optimization algorithm is used to find the optimal parameter and  then final classification is done to identify real or fake using extreme learning classifier model. This model has distinguishing deepfake identification which shows an accuracy of 96.13% on Celeb-V1 dataset compared to existing methods such as MLP-CNN and Yolo InceptionResNetV2 XGBOOST.