Deepfake Detection via Combining Channel and Spatial Attention

Alperen Enes Bayar, Cihan Topal

Abstract: Today, as the widespread use of deepfake technologies weakens the credibility of digital media content, deepfake detection of digital content has become an important issue. Detection of fake content is critical in order to prevent the risk of disinformation that may occur with the rapid spread of manipulated content produced with this technology over the internet. This study proposes a neural network that uses channel and spatial attention mechanisms for the detection of deepfake images. This proposed network is trained with a common dataset by combining DeepfakeTIMIT and VidTIMIT datasets. Compared with models such as InceptionV3, ResNet50 and VGG19, higher accuracy, precision, recall and F1 scores were obtained. This network with attention mechanisms has classified the detection of deepfake images with up to 99% success. The findings of this study will provide an important step in the detection of deep forged images and offer a potential solution for a wider range of applications.

External IDs:dblp:conf/siu/BayarT23