🤖 AI

Beyond the QBER Threshold: A Temporal QBER Based Machine Learning Framework for Multi Attack Detection in BB84 QKD

This paper proposes a temporal QBER-based machine learning framework that utilizes physics-informed features to significantly outperform conventional fixed-threshold monitoring in detecting and classifying stealthy eavesdropping attacks in BB84 Quantum Key Distribution systems.

Original authors: Isha, Deepak Singh, Devesh Kumar, S. K Pal, Praful Hambarde, Amit Shukla

Published 2026-08-06

📖 5 min read🧠 Deep dive

CC0 1.0

Original authors: Isha, Deepak Singh, Devesh Kumar, S. K Pal, Praful Hambarde, Amit Shukla

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ 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

Imagine you are trying to send a secret message using a special kind of flashlight that can only shine in four specific colors. This is the world of Quantum Key Distribution (QKD), a high-tech way for two people to share a secret code that is theoretically unbreakable. The magic trick here is that if anyone tries to peek at the message while it's traveling, the laws of physics say the light itself gets messed up. It's like trying to read a letter written on a soap bubble; the moment you touch it, it pops or changes shape.

In the most popular version of this system, called BB84, the two friends (let's call them Alice and Bob) have a simple rule to check for spies: they count how many times the message got garbled. This count is called the Quantum Bit Error Rate, or QBER. If the error rate goes above a specific "danger line"—set at 11%—they know a spy is there and stop talking. It's like a security guard who only sounds the alarm if a thief trips a laser beam that is set to trigger when someone runs too fast. But what if the thief is a ninja? What if they move so slowly and quietly that they never trip the laser, yet they still manage to steal a few secrets? That is the problem this paper tackles: how do you catch the ninjas who stay just below the alarm line?

The researchers at IIT Mandi and DRDO decided to stop looking at just the average number of errors and started watching the rhythm of the errors. They built a smart computer system, a "temporal QBER-based machine learning framework," that acts like a detective who doesn't just count the broken bubbles but listens to the sound they make when they pop. Instead of waiting for the error rate to hit that scary 11% mark, this new system looks at how the errors happen over time. Do they come in sudden bursts? Do they happen more often when the light is tilted one way? Does the pattern look like a heartbeat or a stutter?

To train their detective, the team created a massive digital simulation of a secret message being sent. They didn't just test one type of spy; they simulated seven different kinds of sneaky attacks, ranging from "Intercept-Resend" (where the spy grabs the message and sends a fake one) to "Trojan Horse" (where the spy sneaks in a hidden tool). They also included a "normal" scenario where no one is spying, just to make sure the system doesn't get paranoid. They fed this data into three different types of smart algorithms: Random Forest, XGBoost, and SVM-RBF. Think of these as three different detectives with different ways of thinking. Random Forest is like a committee of experts voting on the answer; XGBoost is a super-learner that gets better by correcting its own mistakes; and SVM-RBF is a mathematician who draws complex lines to separate the good guys from the bad guys.

The results were a game-changer for catching the ninjas. When they used the old-fashioned method—just checking if the error rate was above 11%—the system was terrible at spotting the sneaky attacks. In fact, it missed the bad guys 84.77% of the time (a False Negative Rate of 0.8477) and only got the right answer 25.82% of the time. It was like a guard who only wakes up if a burglar breaks a window, completely missing the one who picks the lock.

However, when they turned on their new machine learning framework, the story changed dramatically. The best detective, XGBoost, correctly identified the type of attack (or confirmed it was safe) 88.01% of the time. More importantly, it missed the sneaky spies only 1.98% of the time. That is a huge improvement! The system could tell the difference between a "Beam Splitting" attack and a "Time-Shift" attack, even though they look very similar to the old method.

To make sure this wasn't just a lucky guess, the researchers ran the simulation ten times, and the results stayed consistent, proving the system is reliable. They also used a tool called SHAP to ask the computer, "Why did you think that was a spy?" The computer pointed out that it wasn't just the total number of errors that mattered, but specific patterns like how the errors jumped around (burst behavior) and how they changed depending on the angle of the light (basis-dependent asymmetry). These "physics-informed" clues were the key to catching the ninjas.

In short, this paper suggests that by listening to the tempo of the errors rather than just counting them, we can build a much smarter security system for quantum messages. While this work was done in a computer simulation and not yet on a real-world quantum network, it shows a very promising path forward. It proves that we don't need to wait for a disaster (hitting the 11% error limit) to know we are being watched; we can spot the subtle, quiet footsteps of a spy long before they get too close.

Technical Summary: Beyond the QBER Threshold

Problem Statement
Conventional BB84 Quantum Key Distribution (QKD) systems rely on a fixed Quantum Bit Error Rate (QBER) threshold, typically set at approximately 11%, to detect eavesdropping. While effective against aggressive attacks, this static monitoring approach is unreliable in realistic environments characterized by detector imperfections, channel noise, and optical losses. In such conditions, adversaries can execute "stealthy" partial attacks (e.g., intercepting only a fraction of photons) that induce a QBER below the detection threshold while still compromising security. Furthermore, existing machine learning approaches for QKD security often focus on binary detection (attack vs. normal), limited attack scenarios, or scalar QBER features, failing to provide the fine-grained, multi-class identification necessary for effective incident response.

Methodology
The paper proposes a Temporal QBER-based Machine Learning Framework designed to detect and classify seven distinct eavesdropping attacks alongside normal channel operation. The methodology operates as a software-based monitoring layer that utilizes standard BB84 post-processing data without requiring additional hardware modifications.

Temporal Feature Extraction: Instead of relying on average session-level QBER, the framework divides the sifted key into non-overlapping temporal windows (size $W=50$ bits) to generate a temporal QBER sequence. From this sequence, 63 physics-informed features are extracted, categorized into five groups:
- Statistical Features: Mean, variance, skewness, kurtosis, etc.
- Burst & Instability Features: Spike density, burst duration, jump energy, and temporal drift.
- Spectral & Temporal Features: Autocorrelation, Discrete Fourier Transform (DFT) energy, and spectral entropy.
- Basis-Dependent Features: Rectilinear vs. diagonal QBER and basis asymmetry.
- Channel Interaction Features: Key generation rate, transmission efficiency, and QBER-loss interactions.
Classification Models: The extracted feature vectors are fed into three machine learning classifiers: Random Forest (RF), XGBoost, and Support Vector Machine with a Radial Basis Function kernel (SVM-RBF). The system performs multi-class classification to distinguish between a normal channel and seven specific attack types: Intercept-Resend, Photon Number Splitting (PNS), Beam Splitting, Time-Shift, Trojan Horse, Quantum Cloning, and Fake State.
Evaluation Setup: Experiments were conducted on a simulated dataset of 24,000 sessions (3,000 per class) under noisy and lossy conditions. The models were evaluated using 10-fold independent runs with 80/20 stratified train/test splits. Performance was measured against a conventional fixed 11% QBER threshold and a Cumulative Sum (CUSUM) detector.

Key Results

Classification Performance: XGBoost achieved the best overall performance with an average accuracy of 88.01% (±0.47%) and a macro-F1 score of 0.8803. SVM-RBF performed comparably (87.3% accuracy), while Random Forest lagged at 75.7%.
Stealthy Attack Detection: When evaluated as a binary detector (Attack vs. Normal), the conventional 11% QBER threshold achieved only 25.82% accuracy with a False Negative Rate (FNR) of 0.8477. In contrast, the proposed framework reduced the FNR to 0.0198, demonstrating a substantial improvement in detecting attacks that evade threshold-based monitoring.
Per-Class Analysis: The model showed high discrimination for most attacks, with Quantum Cloning achieving an F1 score of 0.999. The primary confusion occurred between PNS and Beam Splitting attacks, and between Time-Shift and Trojan Horse attacks, due to their similar temporal QBER characteristics.
Explainability: SHAP (SHapley Additive exPlanations) analysis revealed that Key generation rate, QBER stability, diagonal-basis QBER, and basis QBER differences were the most influential features, confirming that temporal and channel-aware characteristics are more discriminative than aggregate QBER.
Feature Importance: An ablation study showed that removing "Channel Interaction" features caused the largest accuracy drop (17.19%), followed by "Basis-Dependent" features (3.20%), highlighting the critical role of these physics-informed features.

Significance and Claims
The paper claims that this framework provides an accurate, explainable, and practical solution for multi-attack security monitoring in BB84 QKD systems. Its primary significance lies in moving beyond the limitations of fixed-threshold monitoring and binary detection. By leveraging 63 physics-informed temporal features, the system can identify specific eavesdropping strategies that remain invisible to conventional methods. The authors emphasize that this approach complements existing information-theoretic security guarantees without altering the underlying BB84 protocol, offering a robust software layer for detecting stealthy attacks in realistic, noisy quantum communication environments.

Limitations and Future Work
The study acknowledges that current results are based on simulated BB84 environments with balanced class distributions and simplified attack models. The authors state that future work will focus on experimental validation using practical QKD systems, addressing class-imbalanced and domain-shift scenarios, and developing online adaptive detection for real-time monitoring.

Technical Summary: Beyond the QBER Threshold

More like this