Background. Stress can profoundly affect human behavior. Critical-infrastructure operators (e.g., at nuclear power plants) may make more errors when overstressed; malicious insiders may experience stress while engaging in rogue behavior; and chronic stress has deleterious effects on mental and physical health. If stress could be detected unobtrusively, without requiring special equipment, remedies to these situations could be undertaken. In this study a common computer keyboard and everyday typing are the primary instruments for detecting stress. Aim. The goal of this dissertation is to detect stress via keystroke dynamics – the analysis of a user’s typing rhythms – and to detect the changes to those rhythms concomitant with stress. Additionally, we pinpoint markers for stress (e.g., a 10% increase in typing speed), analogous to the antigens used as markers for blood type. We seek markers that are universal across all typists, as well as markers that apply only to groups or clusters of typists, or even only to individual typists. Data. Five types of data were collected from 116 subjects: (1) demographic data, which can reveal factors (e.g., gender) that influence subjects’ reactions to stress; (2) psychological data, which capture a subject’s general susceptibility to stress and anxiety, as well as his/her current stress state; (3) physiological data (e.g., heart-rate variability and blood pressure) that permit an objective and independent assessment of a subject’s stress level; (4) self-report data, consisting of subjective self-reports regarding the subject’s stress, anxiety, and workload levels; and (5) typing data from subjects, in both neutral and stressed states, measured in terms of keystroke timings – hold and latency times – and typographical errors. Differences in typing rhythms between neutral and stressed states were examined to seek specific markers for stress. Method. An ABA, single-subject design was used, in which subjects act as their own controls. Each subject provided 80 typing samples in each of three conditions: (A) baseline/neutral, (B) induced stress, and (A) post-stress return/recovery-to-baseline. Physiological measures were analyzed to ascertain the subject’s stress level when providing each sample. Typing data were analyzed, using a variety of statistical and machine learning techniques, to elucidate markers of stress. Clustering techniques (e.g., K-means) were also employed to detect groups of users whose responses to stress are similar. Results. Our stressor paradigm was effective for all 116 subjects, as confirmed through analysis of physiological and self-report data. We were able to identify markers for stress within each subject; i.e., we can discriminate between neutral and stressed typing when examining any subject individually. However, despite our best attempts, and the use of state-of-the-art machine learning techniques, we were not able to identify universal markers for stress, across subjects, nor were we able to identify clusters of subjects whose stress responses were similar. Subjects’ stress responses, in typing data, appear to be highly individualized. Consequently, effective deployment in a realworld environment may require an approach similar to that taken in personalized medicine.

Computer security is becoming an increasingly important topic in today’s society, withever increasing connectivity between devices and services. Stolen passwords have thepotential to cause severe damage to companies and individuals alike, leading to therequirement that the security system must be able to detect and prevent fraudulentlogin. Keystroke biometrics is the study of the typing behavior in order to identifythe typist, using features extracted during typing. The features traditionally used inkeystroke biometrics are linear combinations of the timestamps of the keystrokes.This work focuses on feature learning methods and is based on the Carnegie Mellonkeystroke data set. The aim is to investigate if other feature extraction methods canenable improved classification of users. Two methods are employed to extract latentfeatures in the data: Pre-training of an artificial neural network classifier and an autoencoder. Several tests are devised to test the impact of pre-training and compare theresults of a similar network without pre-training. The effect of feature extraction withan autoencoder on a classifier trained on the autoencoder features in combination withthe conventional features is investigated.Using pre-training, I find that the classification accuracy does not improve when using an adaptive learning rate optimizer. However, when a stochastic gradient descentoptimizer is used the accuracy improves by about 8%. Used in conjunction with theconventional features, the features extracted with an autoencoder improve the accuracyof the classifier with about 2%. However, a classifier based on the autoencoder featuresalone is not better than a classifier based on conventional features.