Privacy Preserving Outlier Detection (Wintersemester 2021/2022)
Dozent: Dr. Anne Kayem
(Internet-Technologien und -Systeme)
- Semesterwochenstunden: 4
- ECTS: 6
- Einschreibefrist: 01.10.2021 - 22.10.2021
- Lehrform: Seminar / Übung
- Belegungsart: Wahlpflichtmodul
- Lehrsprache: Englisch
- Maximale Teilnehmerzahl: 10
Studiengänge, Modulgruppen & Module
In an increasingly interconnected world in which almost every device is essentially both a data generator and collector, composing large datasets of complex personal information is ever more easy to achieve. This is in spite of the fact that privacy legislation such as GDPR provides measures to prohibit the collection and storage of personal data without explicit user consent. A further point of alarm is the growing number of reports in popular media on de-anonymization incidents that have paved the way for related security subversion incidents such as leaks of personal data.
In this seminar, we study several anonymized datasets in effort to understand why and how de-anonymizations occur. Specifically, we focus on designing reverse-clustering algorithms to discover outlier data points, and determine how these can be used either individually or in combination with auxiliary data, to de-anonymize data points within the original dataset. As a final point, we will discuss the properties of the outlier data points in terms of how they enabled the de-anonymizations and what possible counter-measures to apply.
Topics to be discussed will include the following:
- Outlier Detection
- Distance-Based Outlier Detection
- Clustering-Based Outlier Detection Approaches
- Model-Based Outlier Detection Approaches
- Algorithms for Outlier Detection
- Ensemble Methods
- Time Series Data ...
There are no pre-requisites for this course, however a background in data mining or machine learning might be helpful.
Reference materials will be provided on a per-lecture and per-need basis.
Lern- und Lehrformen
At the end of this seminar you should have some insight into the research field of outlier detction (or also sometime refered to as anomaly detection) algorithms for supporting the generation of privacy preserving datasets. You will also have studied the conceptual foundations of these algorithms and through the project work applied these learnings to some examples of datasets drawn from real-life practical application areas (e.g. data from COVID tracing apps).
Grading for the seminar will be based on a mid-semester presentation (20%), a final presentation (30%) and a report (50%). The table below provides a summary:
| ||Number ||When ||Grade |
|Mid Semester Presentation ||Group Size (2-3) || ||20% |
|Final Semester Presentation ||Group Size (2-3) || ||30% |
|Final Report ||One (1) per Group || ||50% |