TypeNet: Deep Learning Keystroke Biometrics

We study the performance of Long Short-Term Memory networks for keystroke biometric authentication at large scale in free-text scenarios. For this we explore (LSTMs) trained with a moderate number keystrokes per identity and evaluated under different scenarios including: i) three learning approaches depending on loss function (softmax, contrastive, triplet loss); ii) training samples lengths sequences; iii) four databases based two device types (physical vs touchscreen keyboard); iv) comparison existing both traditional statistical methods deep architectures. Our approach called TypeNet achieves state-of-the-art an Equal Error Rate 2.2% 9.2% physical keyboards, respectively, significantly outperforming previous approaches. experiments demonstrate increase error up to 100,000 subjects, demonstrating potential operate Internet scale. To best our knowledge, used work are largest available research more than 136 million from 168,000 subjects 60,000 63 acquired mobile touchscreens.