Course Overview

The course is divided into two parts that build on each other and last 75 minutes each. The first part includes:

Introduction & Foundations. We introduce the overall topic of authentication, covering aspects of user recognition such as behavioral biometric modalities and the modes of verification and identification, which are used to secure information and enable which personalization of computing devices. Here, we establish a relationship with intelligent user interfaces. Additionally, quality dimensions such as implicitness and continuity and their relevance for usability and user experience are outlined and discussed.

Evaluation & Metrics. We present insights into how to evaluate a novel academic work and discuss the associated metrics (e.g., FAR, FRR, EER, F1-Score) among balanced and unbalanced datasets. We also discuss pitfalls to be aware of when designing a behavioral biometric authentication scheme. Additionally, the issues connected to open-set and closed-set identification are highlighted, and their implications on the employed machine learning models.

Academic perspective. We conclude the first part with insights from a survey conducted in 2022 among the scientific community on the key questions, what is most commonly valued and criticized in novel academic work. We also highlight indicators of rigor. Examples include study design and variables, the reported metrics, and an appropriate split of training, testing and validation data.

The second part of the course is the hands-on session. We provide an application that enables participants to capture behavioral biometric data on stand-alone virtual reality head-mounted displays [7, 8]. Participants’ data is made available to participants in real time through a shared database. We will provide approximately 15 Meta Quest 2 devices with the application installed so that participants can elicit their own data that they will then explore and build an authentication system around, which is then evaluated.

Data elicitation. In the beginning, we will give participants the opportunity to capture their own behavioral biometric data on a virtual reality head-mounted display that we provide. Participation in this step is voluntary as we will additionally provide a backup dataset that is supported by a video of the elicitation.

Data Analytics & Visualization. Next, we provide Jupyter notebook templates in an easy-to-use environment (e.g., Google Colab or Jupyter) to participants that fetch data from the shared database. We introduce participants to common data exploration techniques, such as descriptive statistics on features and visualization techniques. Here, we use Python libraries such as pandas, matplotlib, and scikit-learn. We will focus on finding anomalies within the data that might result in side effects for evaluating a behavioral biometric system.

Machine Learning & Evaluation Metrics. Finally, we will guide participants in creating a machine-learning algorithm to enable the authentication process. We will focus on algorithms that are quick in execution (e.g., Random Forest) and are explainable to a certain degree. In the last step, we will show how to generate the standard metrics to be reported on in an academic publication, as well as any relevant previous preprocessing steps (e.g., trimming and normalization).

Course Format. The course is planned as an in-person event at CHI 2023; however, the first session of the course can be moved to a virtual space. This session would likely take place before the conference and cover the theoretical foundations of behavioral biometrics.