Normalization Technique and Weight Adjustment Analysis for Keystroke Vector Dissimilarity Authentication

A keystroke dynamics authentication uses keystroke rhythm for each user on a keyboard to verify a real user. The idea is that each user has a unique keystroke rhythm such that it can be determined the identity of a user. To verify a user, a keystroke vector dissimilarity technique was proposed to use keystroke features as a vector and calculate a weight using SoftMax+1 to overcome the Euclidean distance problem. However, the weight has yet to be analyzed in detail. Therefore, this paper aims to find a normalization technique and a weight adjustment to enhance the accuracy of the keystroke vector dissimilarity technique. The normalization techniques and activation functions analyzed in this study are Euclidean norm, Mean normalization, Min-max normalization, Z-score normalization, SoftMax function, and ReLU function. The weight adjustment varies from w+1000 to 1000-w. The results show that the Mean and Min-max normalizations with 10-w as a weight gave the same results at 96.97% accuracy and 3.03% error, which are better than the previous work.