The company Bumble, which develops one of the largest online dating services, has opened the source code for the Private Detector machine learning system used to detect obscene images in photos uploaded to the service. The system is written in Python, uses the Tensorflow framework and is distributed under the Apache-2.0 license. For classification, the convolutional neural network EfficientNet v2 is used. A ready-made model for identifying images of naked people is available for download. The determination accuracy is more than 98%.
It also includes a script to create custom models that can be trained on your collection and used to classify arbitrary content. For training, it is enough to run the script with text files containing lists of images with positive and negative properties. After the training is completed, you can transfer an arbitrary image to the Private Detector and, based on it, the hit ratio will be calculated.
Source: opennet.ru