Spectral Analysis of Keystroke Streams: Towards Effective Real-time Continuous User Authentication

Abdullah Alshehri, Frans Coenen, Danushka BollegalaIcissp 2018 · DOI

TL;DR

This paper introduces a novel method, KCASA, for continuous user authentication by analyzing keystroke dynamics in the spectral domain. By transforming keystroke timing data into frequency representations (DFT and DWT) and comparing them using Dynamic Time Warping, KCASA achieves high accuracy, outperforming traditional feature vector methods.

What Problem Does This Paper Address?

Continuous authentication is crucial for applications like online learning platforms where users need to be monitored constantly to ensure their identity. Traditional keystroke dynamics authentication relies on feature vectors derived from timing patterns (like key hold and flight times). However, these methods often require large feature sets, making them computationally expensive and less suitable for real-time continuous authentication. The paper addresses the challenge of developing an efficient and accurate continuous authentication system that can handle the dynamic nature of free-text typing.

Methods

The core of the proposed approach, Keystroke Continuous Authentication based Spectral Analysis (KCASA), involves transforming raw keystroke timing data into the frequency domain. Instead of directly using raw time series data (KTS), the authors explore two spectral transformations:

  • Discrete Fourier Transform (DFT): Converts the time-series data into a sum of sinusoidal functions, representing the data in the frequency domain.
  • Discrete Wavelet Transform (DWT): Another method for time-series transformation, particularly effective for capturing sudden changes and retaining more information than DFT in some cases. The paper specifically uses the Haar wavelet.

After transformation, the similarity between these spectral representations is measured using Dynamic Time Warping (DTW). DTW is chosen over simpler methods like Euclidean distance because it can account for phase shifts and amplitude differences in the time series, making it more robust for comparing typing patterns. The KCASA model operates by continuously sampling subsequences of keystrokes (windows) and comparing them to previously established user profiles. A preprocessing step is also included to reduce noise by capping excessive flight times.

Datasets

The evaluation was conducted using three publicly available datasets:

  • GP (Gunetti and Picardi, 2005): 31 subjects, free text in Italian, using only flight time (Ft) as a feature.
  • VHHS (Vural et al., 2014): Subjects typed free text in English in a lab setting, using both flight time (Ft) and key hold time (KH) as features.
  • ACB (Alshehri et al., 2016b): 30 subjects, free text in English, simulating an eLearning environment, using both flight time (Ft) and key hold time (KH) as features.

Key Insights and Ideas

  • Spectral Domain Advantage: Transforming keystroke dynamics into the spectral domain (using DFT or DWT) can lead to more efficient and accurate pattern recognition compared to raw time-series analysis.
  • DTW's Robustness: Dynamic Time Warping is a crucial component for comparing spectral representations, as it effectively handles variations in typing speed and rhythm.
  • DWT Outperforms DFT for Accuracy: While both spectral methods showed improvement over raw KTS, the DWT representation, particularly when combined with flight time (Ft), yielded the highest authentication accuracy.
  • Parameter Sensitivity: The window size (w) for sampling keystroke subsequences significantly impacts accuracy, with a window of 75 keystrokes generally performing best across datasets and features. The maximum flight time threshold (φ) had a less pronounced effect.
  • Efficiency Gains: The spectral approach (KCASA) offers significant runtime advantages over traditional Feature Vector Representation (FVR) methods, making it more suitable for real-time applications.

Key Results

The proposed KCASA model, particularly with the DWT representation and flight time (Ft) feature, achieved a best overall accuracy of 98.24%. This significantly outperformed the Feature Vector Representation (FVR) approach, which achieved a maximum accuracy of 90.15%. The DWT representation also showed superior performance in terms of False Acceptance Rate (FAR) and False Rejection Rate (FRR) compared to DFT and KTS. While DFT was slightly more efficient than DWT, DWT provided better accuracy.

Limitations and Future Work

  • Parameter Optimization: While the paper explored the effect of window size (w) and flight time threshold (φ), further optimization of these parameters could potentially yield even better results.
  • Multivariate Analysis: The current study primarily focused on univariate time series (either Ft or KH). The authors suggest future work to investigate the use of multivariate keystroke time series, incorporating both Ft and KH simultaneously, to potentially enhance authentication effectiveness.
  • DTW Complexity: The computational complexity of DTW, while manageable, remains an area for potential optimization, especially for very large datasets or real-time systems with strict latency requirements.
  • Dataset Limitations: The GP dataset only provided flight time data, limiting the evaluation of key hold time features on that specific dataset.

Summary updated

BibTeX
@inproceedings{alshehri2018spectral,
  author = {Abdullah Alshehri and Frans Coenen and Danushka Bollegala},
  title = {{Spectral Analysis of Keystroke Streams: Towards Effective Real-time Continuous User Authentication}},
  booktitle = {Icissp},
  year = {2018},
  doi = {10.5220/0006606100620073},
}
Figure