TL;DR
This paper introduces a novel Recurrent Neural Network (RNN)-based approach for user authentication using keystroke dynamics, specifically focusing on freely typed text. By leveraging Long Short-Term Memory (LSTM) networks, their method achieves significantly lower Equal Error Rates (EERs) compared to existing methods, demonstrating its potential for more robust and efficient continuous user authentication.
What Problem Does This Paper Address?
In our increasingly connected world, robust user authentication is paramount. While traditional password-based systems are common, they are vulnerable to compromise. Biometric authentication offers a more secure alternative, but many methods require specialized hardware. Keystroke dynamics (KDA) presents a compelling software-only solution, analyzing typing patterns to verify user identity. However, existing KDA methods struggle with the variability and length of freely typed text, leading to lower authentication performance compared to systems using short, fixed-length inputs. This paper addresses this gap by proposing an RNN-based KDA model that can effectively handle the complexities of long, free-form typing.
Methods
The core of this research lies in a Recurrent Neural Network (RNN) model, specifically employing Long Short-Term Memory (LSTM) cells. The authors chose LSTM due to its proven ability to handle long-range dependencies, a crucial aspect of capturing nuanced typing behaviors.
The proposed RNN-KDA model takes a sequence of two consecutively typed keys (represented as concatenated one-hot vectors) as input. The model is trained to predict four specific keystroke timing values: DD-time (time between pressing two keys), UD-time (time between releasing one key and pressing the next), UU-time (time between releasing two keys), and DU-time (time between pressing and releasing a key).
Crucially, the model is trained per user, learning their unique typing patterns. When a new keystroke sequence is input, the model predicts the expected timing values. The difference between these predicted values and the actual observed timing values forms a "novelty score." A significant deviation indicates a potential impostor. A key advantage highlighted is that this model can be trained solely on a valid user's data, eliminating the need for impostor data during training, which is a common limitation in other machine learning-based KDA methods.
Datasets
The study utilized a dataset collected from 120 participants using a custom web-based keystroke data collector. This collector captured keystroke events, including key identifiers and precise timestamps. The raw keystroke data was then preprocessed to extract five types of timing values, from which four (DD, UD, UU, DU) were used as targets for the RNN model. The study also involved comparing their proposed method against several benchmark methods, which implicitly used similar or derived datasets.
Key Insights and Ideas
- RNNs for Free Text KDA: The paper convincingly demonstrates that RNNs, particularly LSTMs, are well-suited for the challenges of analyzing long, freely typed keystroke data, outperforming traditional statistical methods.
- Predicting Timing for Authentication: Instead of directly classifying patterns, the model's ability to predict keystroke timings for digraphs is a novel and effective approach to identify deviations from a user's normal behavior.
- User-Specific Models without Impostors: The ability to train a robust authentication model using only a valid user's data is a significant practical advantage, simplifying deployment and addressing privacy concerns.
- Adaptability and Continuous Learning: The RNN-KDA model can be updated with new data, allowing it to adapt to changes in a user's typing style over time, further enhancing its long-term effectiveness.
Key Results
The proposed RNN-KDA model achieved significantly superior performance compared to benchmark methods.
- Low Equal Error Rate (EER): The RNN-KDA model achieved EERs as low as 5%-6% using only 10 test keystrokes, while other benchmark methods had EERs above 20%.
- Substantial Improvement: The EER reduction of the RNN-KDA model over the K-S statistic (the second-best method) ranged from 68.62% to 91.23%, with statistical significance at the 0.001 level.
- Efficiency with Few Keystrokes: The model demonstrated high detection accuracy (e.g., 93.3%) with as few as two or three words (approximately 10 test keystrokes) after training on 500 keystrokes.
- Scalability: Performance improved steadily and more rapidly with increasing lengths of training keystrokes compared to benchmark methods.
Limitations and Future Work
The authors acknowledge two primary limitations:
- Device Specificity: The current validation was performed solely on PC keyboard data. Future work should extend this to other input devices like soft keyboards on smartphones to ensure broader applicability.
- Language Dependence: The experiment was conducted using Korean text. To achieve true language independence, the model needs to be verified with languages using different alphabetical systems.
Summary updated