EURECOM Seminar: " Selected Problems in Security and Privacy of Data Science "

Professor Florian Kerschbaum -
Digital Security

Date: -
Location: Eurecom

Bio : Florian Kerschbaum is a professor in the David R. Cheriton School of Computer Science at the University of Waterloo (joined in 2017), a member of the CrySP group, and NSERC/RBC chair in data security (since 2019). Before Iheworked as chief research expert at SAP in Karlsruhe (2005 – 2016) and as a software architect at Arxan Technologies in San Francisco (2002 – 2004). He holds a Ph.D. in computer science from the Karlsruhe Institute of Technology (2010) and a master's degree from Purdue University (2000). He served as the inaugural director of the Waterloo Cybersecurity and Privacy Institute (2018 – 2021). He is an ACM Distinguished Scientist (2019), a winner of the Outstanding Young Computer Science Researcher Award from CS-Can/Info-Can (2019) and a winner of the Faculty of Math Golden Jubilee Research Excellence Award (2022). He is interested in security and privacy in the entire data science lifecycle. He extends real-world systems with cryptographic security mechanisms to achieve (some) provable security guarantees. Abstract : Data Science is the process of handling data from collection, preparation, management, analysis to use. It has a rich set of tools and methods, including databases and machine learning. The processed data is, however, often sensitive or private and the tools and methods of data science need adaption to accommodate security or privacy principles. In this talk, I will highlight selected tools and privacy challenges. First, I will present our research on differential privacy and membership inference attacks. A membership inference attacks tries to infer whether a data sample was part of the training data set of a model just from the model itself and the data sample. Our work shows that there are vastly different bounds on the protection by differential privacy if the data set is iid sampled or not. Second, I will present our work on private collection of key-value data. Key-value data is commonly collected as statistics from devices. We developed a new technique - selective multi-party computation - that combines multi-party computation and differential privacy to achieve a favorable trade-off between performance and privacy.