Advancing Digital Forensics: Comparative Analysis of Deep Learning Models for Deepfake Detection
Mohini Chakarverti, Anurag Goswami, Ashima Yadav · 2024
Deepfake has significantly advanced the generation of synthetic media capabilities, allowing for the creation of significantly promising fake visual and audio content. This technology manipulates existing media to replace a person’s face and voice, posing severe ethical and security concerns. Deepfakes are increasingly exploited in various cybercrimes, including identity theft, cyber extortion, and the spread of fake news, with over 95% involving obscene content highlighting the growing threat. Hence, the detection of deepfake content is a growing field among researchers. Moving on the same lines, our study compared the six well-known models (i.e., MesoNet, Xception, VGG19, ResNet101, NAS-Net, and MobileNet). A dataset was used that has both real and fake images which was employed for all models to assess their performance. Our results showed that the VGG19 was the most effective model with 83.84% among all six. The results can be used to further push the development of tools and techniques to fight against the digital media tampering that may be helpful for the cybersecurity industry.