A common approach to automatically identifying child sexual abuse material (CSAM), is that of dividing the whole task into several steps which includes face detection, age estimation, and sexual-content analysis. Although numerous researchers have claimed remarkable outputs, the majority of these models are not public or generalized for actual forensic detection. There is also a lack of evaluation of the models against benchmark datasets. Considering the risks and sensitivities of training models directly on CSAM, a feasible way to build a reliable CSAM classification model is to take advantage of non-CSAM proxy datasets by introducing transfer learning structures in the model. To avoid the domain shifts, discrepancy-based, adversarial-based or reconstruction-based deep transfer learning methods are adopted in different scenarios.
To examine the performance of deep transfer learning models, a dataset which consists of samples collected from two distinct sources are created. In this dataset, source and target domains are constructed from the different sources. In each dataset, samples are labelled into three categories, i.e. pornography, indicative pornography and clean. Images from Google Open Image, Reddit and other public websites are utilized.
To realize the generalization of the pornography image classification model, two popular transfer learning models, i.e. domain adversarial neural networks and joint adaptation networks are being evaluated on the newly-created transfer learning dataset. The accuracy of classification in each class is reported and compared with the well-established Resnet image classification model. Improvements in overall accuracy demonstrate that the transfer learning model can boost performance of CSAM classifiers, given proper parameter selection.