Accurate Continuous and Non-intrusive User Authentication with Multivariate Keystroke Streaming

Abdullah Alshehri, Frans Coenen, Danushka BollegalaKdir 2017 · DOI

TL;DR

This paper introduces a novel method for continuous user authentication using keystroke dynamics, modeling typing patterns as multivariate time series. This approach significantly improves accuracy and efficiency compared to traditional feature vector methods, making real-time, non-intrusive authentication a practical reality for applications like online learning.

What Problem Does This Paper Address?

The rise of online learning platforms (eLearning, MOOCs) has created a critical need for robust user authentication that extends beyond initial login. Traditional methods like passwords are insufficient for ensuring the identity of users throughout an online assessment. Keystroke dynamics, the unique patterns of how individuals type, offer a promising solution for continuous, non-intrusive authentication. However, existing keystroke dynamics methods often rely on computationally expensive feature vector representations, limiting their real-time applicability and sometimes sacrificing accuracy. This paper addresses the challenge of developing an accurate and efficient continuous authentication system that can operate in real-time, even during extended typing sessions.

Methods

The core innovation of this paper is the shift from feature vector representations to a multivariate time series model for keystroke dynamics. Instead of extracting static features from predefined n-grams (like digraphs or trigraphs), the authors propose modeling the entire sequence of keystrokes as a multivariate time series. Each keystroke event is represented by its timing information, specifically flight time (time between key presses) and hold time (duration a key is pressed).

The authentication process involves:

  • Multivariate Time Series Representation: Keystroke timing data is captured as a sequence of events, where each event is a tuple containing flight and hold times.
  • Sliding Window and Subsequence Extraction: A fixed-size sliding window is used to extract short subsequences from the continuous keystroke time series.
  • Dynamic Time Warping (DTW): DTW is employed to measure the similarity between these subsequences. DTW is chosen for its ability to handle non-linear similarities, making it robust to variations in typing speed and rhythm.
  • Noise Reduction for Flight Time: A mechanism is introduced to cap unusually long flight times, which can occur due to pauses or "away from keyboard" events, preventing them from unduly influencing the authentication process.
  • Start-up Authentication: Initially, a user's identity is confirmed by comparing the first extracted subsequence against a stored reference profile.
  • Continuous Authentication: Subsequent subsequences are compared to the immediately preceding subsequence, allowing for real-time monitoring and authentication without needing to compare against a large database of profiles for every check.

Datasets

The authors evaluated their approach using two datasets:

  • ACB Dataset: Collected by the authors from 30 undergraduate, postgraduate, and staff volunteers at their university. This dataset was collected under less controlled, "real-world" conditions, allowing users to use their own keyboards and environments.
  • VHHS Dataset: Obtained from Vural et al. (2014), this dataset comprised free-text samples from 39 subjects and was collected under laboratory conditions.

For evaluation, each subject's record was split into two halves: one for creating reference profiles and the other for simulating typing streams.

Key Insights and Ideas

  • Time Series Outperforms Feature Vectors: The fundamental insight is that modeling keystroke dynamics as a continuous time series, rather than discrete feature vectors, better captures the nuanced and dynamic nature of typing behavior. This leads to significantly improved authentication accuracy.
  • Multivariate is Better Than Univariate: Incorporating multiple timing features (flight and hold times) in the multivariate time series representation yields better results than using a single feature (univariate time series). This suggests that a richer representation of typing behavior is more discriminative.
  • Real-time Efficiency: The proposed method, by comparing consecutive subsequences rather than a new input against a large profile database, achieves a practical speedup (approximately 1:29) compared to traditional feature vector approaches, making real-time authentication feasible.
  • Privacy Preservation: The mechanism is privacy-preserving as it does not require knowledge of the actual keys being pressed, only their timing.

Key Results

  • Superior Accuracy: The multivariate time series approach achieved a best overall accuracy of 98.39% on the ACB dataset and 97.32% on the VHHS dataset. This is a notable improvement over univariate time series (best accuracy 96.54%) and feature vector approaches (best accuracy 89.15%).
  • Efficiency Gains: The time taken to construct reference profiles was significantly reduced compared to feature vector methods. For a window size of 100, the speedup was approximately 1:29.
  • Low Error Rates: The average FAR was 0.036% and the average FRR was 1.936%, indicating a robust system with low rates of false acceptance and false rejection.
  • Impact of Parameters: The window size ($\omega$) had a more significant impact on performance than the flight time threshold ($\phi$).

Limitations and Future Work

  • DTW Complexity: While DTW is effective, its computational complexity for multivariate time series remains an area for future work. The authors aim to optimize DTW for this specific application.
  • Alternative Representations: Investigating other time series representation methods, such as Fourier transforms, for real-time KCA is suggested for future research.
  • User-Defined Parameters: The window size ($\omega$) is user-defined, and further research could explore adaptive methods for selecting this parameter.

Summary updated

BibTeX
@inproceedings{alshehri2017accurate,
  author = {Abdullah Alshehri and Frans Coenen and Danushka Bollegala},
  title = {{Accurate Continuous and Non-intrusive User Authentication with Multivariate Keystroke Streaming}},
  booktitle = {Kdir},
  year = {2017},
  doi = {10.5220/0006497200610070},
}
Figure