NEURAL NETWORK METHOD OF USER BEHAVIOR ANALYSIS BASED ON KEYSTROKE DYNAMICS
Keywords:
Biometric identification, recurrent neural network, convolutional neural network, behavioral analysis, sliding window, keystroke dynamics, information securityAbstract
The reason for most of the information leaks is the compromise of account data. In this regard,
the introduction of additional means of identification and authentication is relevant. To increase
efficiency, these systems are developed using machine learning. The use of neural networks
is currently the most promising approach to improving the security of systems due to their speed
and accuracy. This paper discusses the use of keystroke dynamics to identify authorized users.
Artificial neural networks are used to analyze the dynamics of pressing. In this paper, such characteristics
as the time of pressing a key, the time between keystrokes, the time between releasing
the first key and pressing the second were analyzed. Both convolutional and recurrent neural networks
were used. The primary processing of input data was carried out by a sliding window that
formed data blocks of a certain size. For further processing of already structured data, a onedimensional
convolutional neural network was chosen, since it is well suited for processing data
presented in the form of a sequence. A recurrent neural network, namely the LSTM architecture,
was used to process time dependencies, since it processes variable-length sequences best and is
less susceptible to gradient decay and explosion than others. For experimental verification of the
effectiveness of this technique, the following architectures were implemented: 2xLSTM, 1D SNC +
LSTM, 1D SNC + 2xLSTM. Based on the results of model training, it was revealed that the system
based on the 1D SNC + 2xLSTM architecture with a sliding window size of 50 has the highest
accuracy. The validation accuracy of this architecture was 98.29%. ROC curves were constructed,
which confirmed the effectiveness of this architecture. The F-measure was calculated, which
showed that the highest performance of binary classification is achieved when using the 1D SNC
+ 2xLSTM architecture with a sliding window size of 50 and equal to 99.39%.
