An Anomaly Detection Based Approach for Continuous Authentication with Smartwatch Inertial Sensors

Arash Gholami, Furkan Alaca, Mohammad Zulkernine.

Computers & Security.

Conventional authentication methods protect unattended devices when they are logged out; however, logged-in devices left unattended are vulnerable to unauthorized access. Inactivity timeouts help mitigate this threat; however, long timeouts increase susceptibility to attack, whereas short timeouts hurt usability. In contrast, continuous authentication mitigates this threat by continuously and non-intrusively verifying whether a device is being used by the user who initially logged in. If verified, the user remains logged in; otherwise, the user is logged out.

We design and evaluate a comprehensive data processing pipeline for smartwatch-based continuous authentication using inertial sensor data. We use a Siamese convolutional neural network to learn and extract discriminative features, and one-class classifiers to determine if a user is the account owner. We compare our learned features with handpicked features proposed in prior work; we show that our learned features achieve better equal-error rates (EER) compared to the handpicked features, particularly for shorter-duration time-series windows. We find that learned features are a promising approach to more quickly and accurately detect unauthorized use of devices. This work thus contributes to making smartwatch-based continuous authentication more secure and usable.

Accepted (Coming Soon)