Privatsphäreerhaltende Kontaktkontextschätzung
Klinische Infektiologie und Tropenmedizin
Zusammenfassung der Projektergebnisse
The CoContexts project explored two directions in the intersection of epidemic respiratory diseases and IT-security: first, advancing techniques for understanding in which sequence of contexts most contacts occurs and, second, data acquisition and novel techniques for validation of public coughing sounds data sets via high-quality clinical coughing sounds data sets. While the limited project duration led to a lack published reports, we do advance the research area at this particular intersection of areas of epidemic respiratory diseases and IT-security with several results, which this report partly provides as reports and partly directly describes. For understanding in which sequence of contexts most contacts occur, we developed novel privacypreserving clustering algorithms and privacy-preserving temporal probabilistic relational models (PRMs) learning techniques. For evaluating the clustering techniques, we have built a simulation of traces of contact context data to evaluate privacy-preserving clustering on triples of contexts. The results show that the privacy-preserving clustering algorithm does learn some information, but more work is needed for accurate clustering. The privacy-preserving temporal PRMs learning constitutes a generic construction that will improve with increasingly accurate privacy-preserving clustering algorithms. Details can be found in the attached report. Concerning the data acquisition, we conducted a lengthy acquisition of clinical data in the context of the university hospital UKSH. Due to pandemic restrictions, our data sets is limited to coughing sounds of 61 patients, but we do offer a high-quality for the labels. Several public data sets with noisy labels were collected during the COVID-19 pandemic. Sensitive clinical data can help understand this data and maximize its utility. We develop a novel technique for privacy-preserving validation of public coughing sounds data sets via high-quality clinical coughing sounds data sets. We base our privacy-preserving method on a recent state-of-the-art meta-learning technique. Details can be found in the attached report. Overall the CoContexts project made advances in the intersection of a medical and a computer science field. The work that we conducted in the CoContexts is the foundation of a joint collaboration that we are currently expanding and plan to formulate in a larger project proposal. While the limited duration hindered us to publish our results within the project duration, the results that we achieved constitute significant advances and first steps towards further research that we expect to see published at top conferences.
