Palm sEMG-based user authentication during doorknob rotation using a convolutional neural network

To exploit the growing demand for authenticated ingress exploiting everyday behavior, contactless use, and ubiquitous Internet-of-Things devices, we propose doorknob-rotation authentication using palm surface electromyography (sEMG). Electrodes on the abductor pollicis brevis and abductor digiti minimi were used to sample signals at 1,000 Hz; the signals were denoised with a 60 Hz notch and 20–500 Hz band-pass, then converted to 2D time–frequency spectrograms via continuous wavelet transform. Subsequently, we used a DenseNet161 convolutional neural network (CNN) to learn fine patterns through dense connections. Using data from five participants, we compared our method with 1D CNN, 2D CNN, and ResNet18 using the same pipeline. The proposed model achieved 94.00% test accuracy and an F1-score of 93.99%, while five-fold cross-validation achieved 91.66 ± 2.78% accuracy and an F1-score of 91.64 ± 2.80%. Palm contact enabled on-device verification without wireless pairing, thereby turning the act of opening a door into authentication and improving convenience and security. The results indicate a practical path for everyday-action sEMG-based authentication and suggest its easy extension to other handles and interfaces.