According to Microsoft, artificial intelligence is proving itself in the fight against child pornography
In 2019 that Microsoft wanted to use automatic image recognition against child pornography. As part of a research project, they cooperated with the state of NRW and the Central and Contact Point Cybercrime NRW. The project is now complete, and the results are available. One is satisfied because the artificial intelligence is to be transferred to productive operation at the ZAC NRW.
A legally compliant and hybrid cloud solution for the automated detection and categorization of child and youth pornography was developed, which has proven itself in the previous test phase in preserving evidence. In 92 percent of all cases, according to Microsoft, it was able to categorize images correctly. After completion of the practical tests, there is now nothing technically to prevent regular operation. Thanks to its open architecture, the cloud-based solution can be used by national and international authorities.
Microsoft advertises its solution as flexible because the AI-based analysis of the suspected material can be carried out both locally in the data centers of the authorities (on-premises) and the Kenyan data centers of Microsoft. As a result, the performance of very large amounts of data can also be expanded at short notice.
Compliance with the legal provisions is ensured by an abstraction algorithm that abstracts and deconstructs (deconstructs) the image files completely and irreversibly. For the human eye, after this deconstruction, no image content is recognizable. This process takes place exclusively in the data centers of law enforcement agencies. The AI algorithm pre-sorts the deconstructed image files. It classifies them according to four categories: whether the images show criminal representations of the abuse of children or adolescents, whether it is permitted adult pornography, or other images.
The investigators can subsequently correct incorrect results. One advantage of the centralized AI solution is that the AI is constantly trained and learns through manual corrections. This should increase the quality in the long run. In addition to the classification, printed or handwritten texts can also be recognized directly on the images via optical character recognition (OCR) and automatically compared with predefined keyword lists. This makes it possible to recognize watermarks, for example, which some perpetrators use. However, this can also help in the evaluation of perpetrator communications