Novel 59-layer dense inception network for robust deepfake identification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Digital Forensics

Background:

• The proliferation of Artificial Intelligence (AI) has enabled advanced media editing tools.
• These tools facilitate the creation and dissemination of deepfakes, sophisticated fake audio and video content used for misinformation and harassment.
• Existing deepfake detection methods often overlook the specific challenges posed by social media platforms.

Purpose of the Study:

• To present a novel deep learning model, FDINet59 (59-Layer Fake Dense Inception Network), for detecting deepfake content.
• To evaluate the efficacy of FDINet59 in identifying deepfakes, particularly those prevalent on social media.
• To assess the model's performance against deepfakes generated by autoencoders and Generative Adversarial Networks (GANs).

Main Methods:

• Development of the 59-Layer Fake Dense Inception Network (FDINet59).
• Training the FDINet59 model using a dataset generated by Multi Task Cascaded Convolutional Networks (MTCNN) cropping.
• Evaluation of the model's deepfake detection capabilities on various datasets, including those generated by autoencoders and GANs.

Main Results:

• FDINet59 achieved a maximum accuracy of 70.02% with a log loss of 0.688 on the training dataset.
• The model demonstrated high performance in detecting deepfakes generated by autoencoders and GANs, reaching 94.95% accuracy with a log loss of 0.205.
• The proposed network shows significant potential for identifying sophisticated fake media.

Conclusions:

• FDINet59 offers a promising solution for detecting deepfake content, especially on social media.
• The model's effectiveness against GAN- and autoencoder-generated deepfakes highlights its robustness.
• Continued development of advanced deepfake detection algorithms is essential to mitigate the societal risks associated with deceptive AI-generated media.