The European General Data Protection Regulation (GDPR) is having an effect in a most unexpected area: autonomous driving. In future, anyone using videos taken from real-world traffic situations for algorithm training will have to comply with special data protection requirements. IAV has therefore developed the IAV Maskin tool to mask data without any loss of information.
The algorithms for autonomous driving need to be trained. The best learning material comes from video clips filmed in real road traffic situations. Real-world scenes let the software learn to recognize pedestrians or other critical situations. Anyone using such clips has to respect the personal rights of the road users featured in the film.
Besides restricting access rights and storage periods, one way of protecting the personal rights of road users featured in films is to mask camera data. Common methods consist in pixelating the faces or using black boxes so that the persons cannot be recognized. However, such clips are practically useless for training algorithms. ”The data then no longer contain important information such as the facial expression and the line of sight“, explains Dr. Mirko Knaak, Product Owner for Artificial Intelligence and Machine Learning at IAV.
IAV is now taking a new approach to ensure it can continue working with video clips. IAV Maskin is the solution that takes personal rights into consideration while also guaranteeing that the images can still be fully analyzed. The system breaks down into several steps. Firstly, it uses artificial intelligence to recognize important facial features in the original data, such as expression and line of vision. It then replaces the originally filmed faces with synthetic equivalents that have the same features. The relevant information is preserved for the algorithms, but it is no longer possible to recognize the filmed persons.
Interaction between ”counterfeit“ and ”police network“
The synthetic faces are created by interaction between two neural networks. The ”counterfeit network“ generates random pictures from white noise, which are assessed by the ”police network“. Initially they have little in common with the original face from the video clip, but millions of iterations result in a synthetic face with the right expression and line of vision as well as the same personality attributes (age, gender, skin color, hairstyle). Once the ”police network“ is satisfied with the similarity, it is automatically superimposed in the video clip. IAV Maskin also takes account of face geometry so that the system even reconstructs movements by the filmed person, such as turning the head.
IAV started to develop IAV Maskin mid-2018, with the first version meanwhile running reliably. ”Our customers are really impressed“, reports Knaak. ”For example, we can generate synthetic faces with the right hairstyles in real time.“ There will probably be great demand for the IAV solution in future, particularly in view of the fact that data protection requirements have been tightened up not just in Europe. Similar regulations also apply in California.
At the CES 2019 in Las Vegas, IAV presented its IAV Maskin tool to the general public with lots of positive feedback. Apart from the automotive industry, many other sectors were also interested in the development, including the American police. The new solution therefore has good chances of giving people a second face all over the world.