Enhancing MobileNet for efficient deepfake detection: a hybrid approach with GhostNet and ShuffleNet

Abstract

Deepfake technology, which involves the manipulation of video, audio, or images using artificial intelligence, has become a growing concern due to its potential for misuse in areas such as information security and social manipulation. Despite advances in deepfake detection, existing models are often too computationally expensive to deploy on mobile or resource-constrained devices. This research aims to address this gap by proposing an enhanced MobileNet architecture tailored for efficient deepfake detection. Through the integration of lightweight modules from GhostNet and ShuffleNet and the application of model compression techniques such as pruning and knowledge distillation, this hybrid model achieves high accuracy with significantly reduced computational complexity. Extensive experiments were conducted to evaluate the model across several deepfake datasets and the results demonstrate that the proposed architecture strikes an optimal balance between accuracy, model size and inference time. These findings provide valuable insights for deploying deepfake detection models in resource-constrained environments, including mobile devices, without compromising performance.