Kaformer: efficient transformer for keystroke authentication

Keystroke dynamics-based authentication leverages an individual’s unique typing behavior for identity verification. However, existing keystroke authentication techniques face two critical challenges: limited recognition accuracy and substantial computational overhead. To address these issues, we propose KAformer, an efficient Transformer architecture for keystroke authentication. Specifically, we introduce a Multi-head Temporal Depthwise Convolutional Attention (MTDCA) mechanism and a Gated Feed-Forward Network (GFFN) to replace the conventional multi-head self-attention and feed-forward network, respectively. These two modules are functionally complementary and tightly coupled. Experimental results on the Aalto desktop and mobile datasets demonstrate that KAformer achieves 99.2% AUC with 0.56% EER, and 99.12% AUC with 0.62% EER, respectively. Furthermore, compared with the standard Transformer architecture, KAformer reduces computational complexity (FLOPs) and the number of parameters by nearly 50%, thereby enhancing operational efficiency while simultaneously reducing computational resource requirements.