An Optimized Privacy Inference Attack Based on Smartwatch Motion Sensors

Wearable devices such as smartwatches have become popular. The accelerometer embedded in smartwatch can record the movement of hand, so that it may lead to privacy compromise when performing sensitive inputting on keyboard with hand wearing smartwatch. The existing inference attack collects smartwatch's accelerometer readings which correspond to the movement of hand when inputting a numeric password, then the inputted password can be inferred by applying classification and inference algorithm to the readings. However, there are two drawbacks in existing inference attack model. Firstly, the sample set is too large, but the attacker is difficult to access large-scale sample data in practice. In addition, the scale of candidate set of possible passwords is sometime too large. Aiming at the two drawbacks of existing inference attack, the optimized attack model proposed in this paper has introduced a correction process and a new inference algorithm. The experiment results show that the optimized attack model can effectively recognize victim's inputted numeric password with small samples and the scale of candidate set is relatively small.