| Abstract [eng] |
Cybersecurity in critical infrastructure requires advanced authentication systems to effectively address the issues of unauthorized access and insider threats. This thesis proposes a novel approach based on the GAFMAT method, which transforms keystroke dynamics data into detailed image representations and thus significantly enhances the ability to distinguish human typing patterns. A Siamese neural network architecture, incorporating convolutional neural network branches, is utilized for the purpose of trustworthy user authentication, thereby enabling the effective distinction between legitimate and illegitimate access attempts. To enhance the accuracy of the authentication process and adapt the methodology to all password lengths, data fusion techniques are employed to standardize the input data from different datasets with different password lengths. Experimental evaluation has shown that the proposed GAFMAT method achieved an equal error rate of 0.04545 in the CMU dataset, indicating that it is considerably outperforming other non-image to image transformation methods. In addition, advanced multidimensional visualization techniques provide support for cybersecurity decision making. The results underscore the effectiveness and practical applicability of the presented approach in enhancing cybersecurity for critical infrastructure. |
