Currently, the implementation of machine learning is mainly about speed and building highly accurate models. Are security aspects being neglected? How important is securing the resulting machine learning systems for machine learning practitioners? These questions are to be answered by a joint research project of the Fraunhofer Institutes for Applied and Integrated Security AISEC and for Secure Information Technology SIT in cooperation with the National Research Center for Applied Cybersecurity ATHENE and Freie Universität Berlin.

Companies are making increasing use of machine learning (ML) processes. They hope to gain competitive advantages, for example by offering their customers new services or optimizing their processes. In order not to waste time, the new tools are often designed and implemented very quickly. This might in many scenarios result in a neglection of security aspects. The threat scenarios are, however, diverse and range from data theft to manipulation of the ML algorithms and training data being used. At the same time, the tightened specifications regarding data protection (e. g. GDPR) are particularly relevant for ML, since the data that is usually used for training the models belongs to the users. However, so far, there is a lack of knowledge about how individual data points influence the underlying model and how easily such data points can be reverse-engineered from model parameters. If companies and their ML practitioners do not take these security aspects seriously, not only economic and financial damage is threatened, but also reputation and image.

There is still little information available on the extent to which security aspects of ML procedures are already considered in practice, and how ML practitioners deal with them. Gaining insights into these questions is the aim of a joint study by the Fraunhofer Institutes AISEC and SIT in cooperation with ATHENE and FU Berlin. The study results will be presented at the beginning of next year.