Exploring Feature Map Correlations for Effective Fake Face Detection
Narges Honarjoo, Fatemeh Taher, Azadeh Mansouri · 2024
In recent years, the internet and social media platforms have seen a massive influx of data, particularly images. To effectively identify fake content, we employ a deepfake detection technique aimed at determining the authenticity of uploaded material. In this paper, we achieve fake face detection with high accuracy by utilizing deep feature map relations. Our approach leverages transfer learning techniques from pretrained deep models, such as VGG16, VGG19, and ResNet50. We analyze the generated Gram Matrix of each layer, where deep feature relations are similar to the covariance matrix, reflecting correlations between different feature maps. These features are used to classify content as fake or real using a simple CNN, and the results are reported and compared. This study explores deep feature map relations as crucial indicators for detecting inconsistencies introduced by image generators. By employing pre-trained networks, we analyze and explore feature map relations generated in each layer as higher-order quality features. Experimental results demonstrate that the proposed inconsistency detection feature yields satisfactory results. The entire implementation of this work can be found at https: //github.com/hnarges91/Fake-Face-Detection.