💻 computer science

Lightweight, Autonomous Security Agents for Operator Workstations: An Explainable XGBoost-Driven Authentication Policy

This paper presents a lightweight, autonomous security agent for operator workstations that utilizes an efficient XGBoost-based keystroke dynamics module to achieve 99.8% authentication accuracy with minimal resource consumption, while employing SHAP-driven explainability and graded, human-in-the-loop policies to enable sustainable, mass-scale edge deployment.

Original authors: Erik Pál, Róbert Lakatos, András Hajdu

Published 2026-10-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Erik Pál, Róbert Lakatos, András Hajdu

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the modern world, digital security often relies on a simple ritual: typing a password and hoping it is enough. But once a user logs in, that single check is forgotten, leaving the session open for anyone who might take over the computer. This is a particular danger in places like power grid control rooms, where an intruder with access to an operator's screen could issue commands that affect the physical world. To solve this, researchers have long looked at how people type. Everyone has a unique rhythm, a specific speed and timing between pressing keys that is as distinctive as a fingerprint. This "typing rhythm" can be measured passively, without any special hardware, simply by watching the flow of keystrokes as a person works. The challenge has been to build a system that can watch this rhythm constantly, spot a stranger instantly, and do so without draining the computer's battery or slowing down the operator's work.

A team of researchers at the University of Debrecen has built a new kind of security guard for computer workstations that operates on these principles. Instead of using heavy, energy-hungry artificial intelligence models that require powerful servers, they created a lightweight, self-contained agent that lives directly on the operator's computer. This agent acts as a continuous observer, watching the typing stream and using a fast, efficient decision-making tool to decide if the person at the keyboard is who they claim to be. If the agent senses a stranger, it does not immediately lock the user out, which could be dangerous during a critical moment. Instead, it follows a careful, graded path of warnings and challenges, always explaining its reasoning in plain language so the human operator can understand and verify the decision.

The core of this system is a clever piece of software that learns the operator's typing style. The researchers trained this software on a dataset of over 13,000 keystrokes from five different people. The system breaks the typing stream into small windows of fifteen keystrokes and analyzes the timing of each key press, the pauses between them, and the overall speed. It then compares this pattern against what it knows about the legitimate user. In their tests, this system was remarkably accurate, correctly identifying the right person 99.8 percent of the time. More importantly, it was incredibly fast and light. It made a decision in just 63 milliseconds and used only about 20 megabytes of memory. To put this in perspective, the researchers compared it to a more complex, deep-learning system that took 180 milliseconds to decide and used more than seven times as much memory. The new system was also vastly more efficient to train, taking just over one second to learn, whereas the older method required nearly ninety seconds.

Because the system is so efficient, it can run on the computer itself without needing to send sensitive data to a central cloud server. This keeps the raw typing information private and reduces the energy needed to run the security check. The researchers designed the agent to be "green," meaning it minimizes the energy cost of keeping the system secure. It only performs the most expensive tasks, like generating a detailed explanation of why it is suspicious, when it actually needs to intervene. If the typing looks normal, the agent stays silent and uses almost no power. When it does detect something wrong, it waits to see if the suspicion is a one-time glitch or a real threat. It requires two or three consecutive suspicious moments before it acts, which prevents it from annoying the user with false alarms.

When the agent does decide to intervene, it follows a clear, human-friendly process. First, it might ask the user to confirm their identity. If the user fails to respond or if the suspicion grows stronger, the agent issues an alert. Only if the situation becomes critical, and the user has ignored previous warnings, does the agent lock the workstation. Crucially, every time the agent raises an alarm, it provides a simple explanation. It tells the operator exactly which part of their typing rhythm looked different, such as "the time you held this key was much longer than usual." This transparency allows the human to understand the decision and correct it if the agent made a mistake. The researchers tested this entire process by replaying recorded sessions of both the real user and four different impostors. The system successfully locked out three of the four impostors within fifteen to twenty-two seconds of them starting to type, while never interrupting the legitimate user.

The study highlights that high security does not have to come at the cost of speed or energy. By using a lightweight, explainable approach, the researchers created a system that is precise enough to catch intruders quickly but gentle enough to avoid disrupting real work. The agent operates with a "human-in-the-loop" philosophy, meaning it makes the initial detection but leaves the final, high-stakes decisions to the human operator, supported by clear evidence. This design ensures that even if the central monitoring system goes offline, the individual workstations remain protected. The results suggest that it is possible to deploy this kind of continuous, intelligent protection across thousands of computers without overwhelming their resources, offering a practical way to secure critical infrastructure against the subtle threat of a hijacked session.

Technical Summary: Lightweight, Autonomous Security Agents for Operator Workstations

Problem Statement
Protecting operator workstations in critical infrastructure, such as electrical power grids and SCADA systems, requires continuous verification of user identity beyond the initial login. Traditional authentication fails to detect session hijacking or takeover by an adversary who gains access to an already-open session. While continuous authentication via keystroke dynamics offers a passive solution, existing deep learning approaches (e.g., LSTM-based TypeNet) are often too resource-intensive for deployment on distributed, resource-constrained edge workstations. Furthermore, autonomous security systems in safety-critical environments must be explainable and provable to ensure human operators can trust and verify automated interventions. The challenge is to design a system that is lightweight, energy-efficient, accurate, and capable of autonomous action while maintaining a "human-in-the-loop" (HITL) for high-risk decisions.

Methodology
The authors propose a lightweight, autonomous security agent that operates locally on the operator's workstation. The system follows a sensory–cognitive–action cycle:

Sensory Module: A custom keylogger captures low-level keyboard events (press/release times, hold times, seek times, error rates) and segments them into non-overlapping windows (15 keystrokes). It engineers a feature vector (609 dimensions) including statistical aggregates, n-gram rhythms, and cumulative typing speeds.
Cognitive Module: Instead of deep sequence models, the agent employs an XGBoost (gradient-boosted decision tree) classifier. This module acts as the agent's "belief-forming" mechanism, outputting a probability that the current typist is an impostor. The model is trained on engineered keystroke features without resampling, distinguishing the enrolled "Main" user from "Other" users.
Decision Policy: The agent does not act on single-window scores. It utilizes a state-dependent policy that requires sustained evidence to trigger interventions. The policy escalates through four levels: Allow, Challenge (requiring 2 consecutive suspicious windows), Alert (requiring 3 consecutive windows above a critical threshold), and Critical (lockout). Thresholds ( $\tau_w$ and $\tau_c$ ) are dynamically derived from the model's Receiver Operating Characteristic (ROC) curve and Equal Error Rate (EER) on a held-out test set.
Action and Explanation: Interventions trigger specific actions (e.g., re-authentication dialogs). Crucially, every intervention is accompanied by a SHAP (SHapley Additive exPlanations) explanation, specifically using TreeSHAP to identify the top 5 timing features that contributed to the impostor score. This explanation is presented to the operator and a supervisor dashboard.
Human-in-the-Loop and Adaptation: The system operates under PEP-AI (Precise, Explainable, Provable AI) principles. High-risk actions (lockouts) are reversible by the operator via OS re-authentication. Model adaptation (retraining) is not automatic; it requires explicit human review and approval to prevent adversarial poisoning.
Energy-Oriented Framework: The authors evaluate the system not just on accuracy but on four energy-related dimensions: compute effort per decision, memory footprint, adaptation cost, and data transmission volume. They compare XGBoost against an LSTM-based TypeNet baseline, using inference time and memory usage as proxies for energy consumption under an equal-power assumption.

Key Contributions

Formal Agent Architecture: A complete framework for keystroke-based continuous authentication that integrates sensing, cognitive processing (XGBoost), and graded action policies into a single autonomous agent.
Graded, Explainable Intervention Policy: A policy that escalates from silent allowance to lockout based on consecutive suspicious windows, with thresholds derived from model error characteristics. It includes a SHAP-based channel that provides human-readable justifications for interventions without revealing raw keystroke data.
Privacy-Preserving Edge Design: A design where raw biometric data and the model remain on the edge device. Only decisions, aggregate timing statistics, and SHAP explanations are transmitted to a supervisor, ensuring privacy and reducing bandwidth.
Energy-Oriented Evaluation: A reinterpretation of standard metrics (latency, memory, training time) as indicators of energy efficiency, demonstrating that a lightweight tree ensemble can outperform deep learning baselines in both accuracy and resource efficiency.

Results
Evaluated on a benchmark of 13,251 keystrokes from 5 users:

Accuracy: The XGBoost module achieved 99.8% accuracy, outperforming the TypeNet baseline (98.6%). The module reduced the Equal Error Rate (EER) from 0.0167% to 0.0017%. In the agent context, this 0.0017% EER is reinterpreted as the keystroke-level False Intervention Rate (FIR), representing the fraction of legitimate operator keystrokes treated as intrusion at the EER operating point. (Note: The reference implementation's specific model version used for threshold derivation showed an EER of 1.68% on its own session-wise test set, but the primary benchmark comparison highlights the 0.0017% EER improvement).
Efficiency: The XGBoost agent demonstrated significantly lower resource consumption:
- Inference Time: 63 ms vs. 180 ms for TypeNet (approx. 2.86x faster).
- Memory: Peak memory usage of 20.52 MB vs. 145.90 MB (approx. 7x smaller).
- Model Size: 0.18 MB vs. 1.99 MB (approx. 11x smaller).
- Training Time: 1.02 s vs. 87.85 s (approx. 86x faster).
End-to-End Performance: In a replay of held-out sessions, the agent successfully locked out three of four impostors within 15–22 seconds of typing. It generated zero false interventions on the enrolled operator's stream, whereas a single-window policy would have triggered two.
Explanation Quality: SHAP explanations were found to be faithful (resetting top features significantly reduced impostor probability), stable (consistent reasons across consecutive windows), and verifiable (96.6% of reported values fell outside the operator's usual range).

Significance and Claims
The paper claims that this design enables "mass edge deployment" of continuous authentication in critical infrastructures by aligning with Green AI principles. The authors argue that security and energy efficiency are not opposing goals in this context; the more accurate model (XGBoost) is also the more computationally efficient one. By localizing computation and decision-making, the system eliminates the need for GPUs, permanent cloud connectivity, and the transmission of raw behavioral data.

The significance lies in the shift from passive prediction to autonomous, explainable action. The system provides a "Precise, Explainable, and Provable" (PEP-AI) security layer where every intervention is justified by data-driven evidence (SHAP) and logged for audit. The authors emphasize that while the agent acts autonomously on low-risk decisions, it retains a human-in-the-loop for high-stakes actions (lockouts and model retraining), ensuring that the system remains trustworthy and adaptable in safety-critical environments like power grid control rooms. The work extends previous conference findings by formalizing the agent architecture, introducing the graded policy, and providing a comprehensive energy-oriented evaluation.

More like this