Enabling privacy-preserving first-person cameras using low-power sensors

Wearable smart devices such as smart-glasses, smart-watches and life-logging devices are becoming increasingly popular, and majority of them are being equipped with first-person cameras. Such first-person cameras on smart-glasses or lifeloggers capture photos/videos from user's point of view, allowing them to record and share user's everyday events. However, these wearable devices with first-person cameras raise serious privacy concerns because they can also capture extremely private moments and sensitive information of the user. Currently, such devices lack the intelligence to understand user's preferences about certain scenarios being sensitive/private. To address this problem, we present PriFir, a scheme that enables Privacy-preserving First-person cameras. PriFir is based on the idea that low-power sensors (e.g. accelerometer, light sensor, etc.) embedded in smartphones and smart-watches can be leveraged to identify sensitive scenarios. Learning from user's preferences, PriFir employs a cascade of classifiers that tags a scenario to be sensitive simply based on the characteristics of the low-power sensor data. We evaluate PriFir using real sensor traces spanning over multiple days and show that it performs highly accurate classification at a low energy cost.