The ISC International Journal of Information Security

Abstract Face manipulation techniques have raised widespread public concern. Although conventional convolutional neural networks (CNNs) demonstrate satisfactory performance, they exhibit limitations in capturing the critical features that are directly affected by manipulation artifacts. Fake faces often exhibit unnatural correlations between feature maps, particularly in local patterns. The Gram matrix can illustrate these irregularities in feature map relationships and support discrimination between real and fake images. In this paper, we show that the relationships among convolutional neural network feature maps typically display natural and coherent patterns; in contrast, manipulated faces tend to disrupt this consistency, which can be effectively captured by analysing the Gram matrix of the feature maps. Experimental results demonstrate that the proposed correlation-based features achieve satisfactory performance in both single-dataset and cross-dataset validation.