Making Activity Recognition Robust against Deceptive Behavior

• Published in: PloS one 2015-12, Vol.10 (12), p.e0144795-e0144795
• Main Authors: Saeb, Sohrab, Körding, Konrad, Mohr, David C
• Format: Article
• Language: English
• Subjects: Activities of daily living; Activity recognition; Adult; Analysis; Classifiers; Deception; Ergometry - instrumentation; Ergometry - standards; Exercise; Female; Health care; Health insurance industry; Humans; Incentives; Machine Learning; Male; Motivation; Motor Activity; Pattern Recognition, Automated - standards; Physiological aspects; Preventive medicine; Retraining; Robustness; Sensors; Services; Shaking; Smartphone - instrumentation; Smartphone - standards; Smartphones; Studies; Tracking devices
• ISSN: 1932-6203; 1932-6203
• DOI: 10.1371/journal.pone.0144795

Healthcare services increasingly use the activity recognition technology to track the daily activities of individuals. In some cases, this is used to provide incentives. For example, some health insurance companies offer discount to customers who are physically active, based on the data collected from their activity tracking devices. Therefore, there is an increasing motivation for individuals to cheat, by making activity trackers detect activities that increase their benefits rather than the ones they actually do. In this study, we used a novel method to make activity recognition robust against deceptive behavior. We asked 14 subjects to attempt to trick our smartphone-based activity classifier by making it detect an activity other than the one they actually performed, for example by shaking the phone while seated to make the classifier detect walking. If they succeeded, we used their motion data to retrain the classifier, and asked them to try to trick it again. The experiment ended when subjects could no longer cheat. We found that some subjects were not able to trick the classifier at all, while others required five rounds of retraining. While classifiers trained on normal activity data predicted true activity with ~38% accuracy, training on the data gathered during the deceptive behavior increased their accuracy to ~84%. We conclude that learning the deceptive behavior of one individual helps to detect the deceptive behavior of others. Thus, we can make current activity recognition robust to deception by including deceptive activity data from a few individuals.