AI Summary & Key Takeaways
AI InsightExperimental results demonstrated that the ResNet-50 model (25.6 M), which contains nearly the same number of parameters as ResNeXt-50—the model with the fewest parameters (25.0 M)—achieved the highest performance among all ResNet models in terms of most confusion metrics, with an accuracy of 0.8633, an F1-score of 0.8441, a precision of 0.7862, and a specificity of 0.8304.
📌 3 Key Takeaways
Abstract
Today, in parallel with technological advancements, the proliferation of image editing tools and artificial intelligence-based manipulation systems has led to an increase in digital image forgery, which has brought about significant problems at the societal level concerning reliability, security, and disinformation. In this context, it is observed that the majority of recent studies aimed at addressing this problem focus on deep learning (DL) and architectures derived therefrom. In this study, the effectiveness of DL-based residual neural network (ResNet) architectures in forgery detection was examined and analyzed through a comparative approach along the axes of depth and structural diversity. Standard-depth models comprising ResNet-50, ResNet-101, and ResNet-152 as ResNet architectures, along with structural variants including Wide ResNet-50 and ResNeXt-50, were trained and tested on the CASIA v2.0 dataset, which is regarded as a benchmark in the image forgery literature. Contrary to the common assumption, the analyses indicate that increasing depth does not always improve performance in forgery detection. Experimental results demonstrated that the ResNet-50 model (25.6 M), which contains nearly the same number of parameters as ResNeXt-50—the model with the fewest parameters (25.0 M)—achieved the highest performance among all ResNet models in terms of most confusion metrics, with an accuracy of 0.8633, an F1-score of 0.8441, a precision of 0.7862, and a specificity of 0.8304. Although the Wide ResNet-50 model performed better in recall and AUC metrics with scores of 0.9151 and 0.9353, respectively, the performance gap is marginal when compared to the corresponding metric scores of the ResNet-50 model. Considering that the Wide ResNet-50 model is significantly more complex than ResNet-50 in terms of the number of parameters, ResNet-50 has emerged as the prominent model among all ResNet architectures in image forgery detection in terms of overall performance. The findings reveal that architectural design choices constitute a more decisive factor than the mere number of layers in image forgery detection.