Electrocardiogram signals have unique characteristics and structure that are difficult to fake. This is due to the fact that the electrical activity of the heart is unique for each person. In addition, the main biometric parameters are hidden during ECG authentication, which makes the process more secure and protected from counterfeiting. The purpose of this study is to evaluate the effectiveness of neural networks for ECG authentication for signals with non-periodic cardiac arrhythmias. The Siamese neural network has been developed as a model. The stages of preprocessing of ECG signals taken from the MIT-BIH database are also described. The model presented in the paper has achieved the following results. Accuracy: 99.69%. Sensitivity: 99.43%. Specificity: 99.94%. ROC-AUC: 99.69%. The results allow us to conclude that the proposed model can effectively authenticate users who have non-periodic cardiac arrhythmias, provided at least a small number of registered standards with violations.

Keywords: biometric authentication, ECG, Siamese neural network, convolutional neural network, Euclidean distance, ROC analysis