Evaluating the effectiveness of rationale-augmented convolutional neural networks for deepfake detection
Küçük Resim Yok
Tarih
2023
Yazarlar
Dergi Başlığı
Dergi ISSN
Cilt Başlığı
Yayıncı
Springer
Erişim Hakkı
info:eu-repo/semantics/closedAccess
Özet
Deepfake image detection has emerged as an important area of research due to its wide-ranging implications for various security systems. In particular, in the field of deep learning, the task of detecting fake images has traditionally been challenging due to its complicated and abstract nature, especially in the field of computer vision where accurate analysis and understanding of facial landmarks play a crucial role. This study introduces a rational-augmented convolutional neural network (RACNN) for deepfake image detection. The RACNN combines a convolutional neural network (CNN) with a reasoning generator, which generates binary masks to highlight the crucial regions that contribute to the CNN's decision-making process. To improve the accuracy and efficiency of the reasoning generator, a reinforcement learning technique is used to train it to generate accurate and compact masks. Through extensive experiments conducted on a large dataset of deepfake images, the effectiveness of the RACNN method is demonstrated, achieving an impressive accuracy rate of 94.87% on an open-source dataset, namely FaceForensics++. The comparative analysis shows the superiority of the RACNN model over existing approaches, especially in terms of accuracy. This robustly demonstrates the effectiveness of the RACNN in accurately distinguishing between real and fake images. The AUC of 95.69% on the dataset serves as a strong indication of the effectiveness of our proposed method in accurately detecting fake facial images generated by various deepfake techniques. Our model proves to be a promising way to advance the field of deepfake image detection, providing potential improvements to the capabilities of such systems.
Açıklama
Anahtar Kelimeler
Deepfake, Convolutional neural network, Rationale-augmented CNN, Forgery detection
Kaynak
Soft Computing
WoS Q Değeri
Q2
Scopus Q Değeri
Q2
