AttM-CNN: Attention and metric learning based CNN for pornography, age and Child Sexual Abuse (CSA) Detection in images

Seeing without Looking: Analysis Pipeline for Child Sexual Abuse Datasets

“…While Rondeau [40] leverages label distributions to assess apparent age, Macedo et al [30] propose a single-model estimation of child presence, age, and gender. A more recent work [18] explores a wide range of technical improvements over neural networks such as residual connections along with inception and attention modules to propose separate models for age estimation and pornography detection.…”

“…Gangwar et al [18] also propose Juvenile-80k, gathering around 24 thousand underage images from a wide range of age estimation datasets and supplementing it with images crawled from public search engines. It is important to note that this type of collection does not abide by ethical standards such as UNICEF's Responsible Data for Children report [52], but it indicates an important gap in the literature: age estimation models specialized in children and adolescents.…”

“…They gathered over 19 thousand faces of individuals under 18 years old, from creative-commons licensed Flickr images. Similar to Gangwar et al [18], Castrillón-Santana et al [7] and Chaves et al [9] gather images from several age estimation databases, amounting to a large number of underage samples, with no extra supplementation of data. Castrillón-Santana et al [7] propose AgeMega, a dataset with over 15 thousand underage faces and around 30 thousand adult ones.…”

“…Apart from skin tone, all features we leverage are commonly estimated from facial images, evoking the need for a face detection module, which by itself generates relevant attributes (refer to Table 1). Many references in the literature rely on MTCNN [54] as the chosen tool for face detection [9,18,30] since it is one of the most accurate and robust in the literature. Therefore, it was the model of our choice as well.…”

“…A comparative evaluation on pornography detection methods [17] found that Yahoo's OpenNSFW model [31] achieved the best accuracy, over 87%, in a CSA database provided by the Spanish Police, despite it being a model trained solely on adult pornography. Gangwar et al [18] found that combining OpenNSFW with an age estimation model achieves over 83% accuracy for CSAM detection in more challenging settings, including adult pornography in the test database. Although their proposed model outperforms OpenNSFW, neither the model nor the dataset used for…”