⚡ electrical engineering

Detection of False Data Injection Attacks in Nuclear Power Plants Using Particle Filters and Time-Series Transformers

This paper proposes a physics-assisted defense framework that integrates non-linear Particle Filters with Time-Series Transformers to effectively detect stealthy False Data Injection Attacks in nuclear power plants by identifying the thermodynamic and temporal decoupling they induce across coupled sensor channels, outperforming conventional linear residual-based detectors.

Original authors: Shashank Padasalagi, Harshitha Pawar

Published 2026-09-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Shashank Padasalagi, Harshitha Pawar

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

Modern cities and industries rely on vast networks of sensors and computers that act as the nervous system for critical infrastructure. In a nuclear power plant, this system constantly monitors the heart of the reactor, tracking how fast the nuclear reaction is happening, how hot the fuel is, and how the cooling water flows. These digital controls, known as cyber-physical systems, allow for precise and safe operation, but they also open a door for digital intruders. The danger is not just a computer crash, but a "false data injection" attack. In this scenario, a hacker does not break the physical machinery; instead, they slip a lie into the data stream. They tell the control room that the temperature is normal when it is actually rising, or that the water flow is steady when it is stopping. If the computer believes this lie, it might fail to take emergency action, potentially leading to a disaster. The challenge for security experts is that these lies can be crafted to look perfectly consistent with the laws of physics, tricking standard alarms that only look for simple mathematical errors.

Researchers Shashank Padasalagi and Harshitha Pawar from M S Ramaiah University of Applied Sciences have developed a new way to catch these sophisticated lies. Their work focuses on the unique, complex relationship between the different parts of a nuclear reactor. They realized that while a hacker might be able to fake a single sensor reading, they cannot easily fake the entire story of how the reactor behaves over time. In a real reactor, if the temperature changes, it forces a chain reaction of other changes: the fuel expands, the neutron flow shifts, and the concentration of certain radioactive byproducts drifts in a specific, predictable way. These relationships are governed by the laws of thermodynamics and nuclear physics, which are non-linear and deeply interconnected. A hacker trying to spoof the data usually relies on a simplified, straight-line model of how the reactor works. This creates a subtle disconnect: the fake data might look right for a split second, but it fails to hold up when viewed as a continuous, evolving story across multiple sensors.

To expose this disconnect, the researchers built a dual-layer defense system. The first layer acts like a rigorous physics check. They used a method called a particle filter, which can be thought of as running thousands of tiny, parallel simulations of the reactor at the same time. Each simulation represents a possible state of the plant, and they are constantly updated with real sensor data. If the incoming data is a lie, these thousands of simulations will start to drift apart from the data, creating a statistical mismatch that a standard computer would miss. The second layer uses a time-series transformer, a type of artificial intelligence designed to understand sequences and patterns over time. This system looks at the history of the sensor readings and the results from the physics check. It pays close attention to how different sensors talk to each other. For instance, it checks if a drop in temperature is accompanied by the expected rise in neutron flow and the correct shift in chemical byproducts. If the data shows a temperature drop without the necessary physical consequences in the other channels, the system flags it as a lie.

The team tested this approach using a detailed computer simulation of a pressurized water reactor. They created a dataset containing over 629,000 time-stamped observations, covering normal operations and various types of attacks. In one scenario, they simulated a "stealth" attack where a hacker slowly manipulated the temperature data to make it look like the reactor was cooling down, even though the core was actually heating up. Standard security systems, which rely on simple linear checks, failed to notice this because the fake numbers fit the basic equations. However, the new dual-layer system caught the deception. It detected that the fake cooling trend was physically impossible given the behavior of the other sensors. The system identified the attack with an accuracy of nearly 99.9 percent, correctly distinguishing between safe operations, obvious errors, and these hidden, sophisticated lies.

The results of the simulation show that the proposed method is highly effective at spotting attacks that traditional tools miss. The system correctly identified every single instance of a clumsy, obvious attack and caught the vast majority of the subtle, stealthy ones, with very few false alarms. By combining a deep understanding of nuclear physics with advanced pattern recognition, the researchers demonstrated that it is possible to see through the deception. The study suggests that while attackers can manipulate data to look consistent with simple rules, they cannot easily mimic the complex, non-linear dance of a real nuclear reactor. This approach offers a promising path forward for securing critical energy infrastructure, ensuring that even if a hacker tries to rewrite the story of what is happening inside a reactor, the truth will eventually be revealed.

Technical Summary: Detection of False Data Injection Attacks in Nuclear Power Plants Using Particle Filters and Time-Series Transformers

1. Problem Statement

The integration of Cyber-Physical Systems (CPS) into nuclear power plants, specifically through digital Instrumentation and Control (I&C) and SCADA networks, has expanded the attack surface for cyber adversaries. A critical vulnerability identified is the False Data Injection Attack (FDIA), where attackers manipulate sensor measurements (e.g., coolant temperature, flow rates) to induce unsafe physical conditions while remaining undetected by traditional monitoring systems.

The paper highlights that conventional detection methods, such as Linear Kalman Filters combined with Chi-squared residual tests, are susceptible to "algebraic stealth attacks." In these scenarios, an adversary with knowledge of the system's linear observation matrix ( $C$ ) injects a malicious vector ( $a_k$ ) that lies within the column space of $C$ . This allows the attacker to bias the state estimates without altering the innovation residuals ( $r_k$ ), effectively rendering linear detectors blind to the manipulation. While these attacks succeed in the discrete linear domain, the paper argues they inherently violate the continuous non-linear thermodynamic and kinetic constraints governing the actual reactor physics.

2. Methodology

The authors propose a Physics-Assisted Defense Framework that exploits the discrepancy between discrete linear stealth assumptions and continuous non-linear reactor dynamics. The framework consists of two primary layers:

A. Theoretical Foundation: Continuous Decoupling

The core theoretical contribution is the proof that while an attacker can satisfy discrete linear constraints ( $c_{k+1} = A c_k$ ), they cannot satisfy the continuous non-linear Point Reactor Kinetics (PRK) equations ( $\dot{c}(t) \neq A_c c(t)$ ).

Theorem 1: Demonstrates that any algebraic stealth injection preserving discrete residual invariance inherently violates the continuous governing equations due to non-linear coupling terms ( $\mathcal{H}$ ) and operating point shifts.
Implication: An attacker cannot spoof a single sensor channel without inducing detectable thermodynamic and kinetic contradictions across coupled states (neutron flux, precursor concentration, fuel temperature, Xenon poisoning, and valve positioning).

B. The Dual-Layer Detection Framework

To detect these violations in real-time, the system employs:

Non-Linear Particle Filter (Statistical Divergence):
- Unlike linear estimators, the Particle Filter propagates a swarm of particles through the non-linear PRK equations.
- It computes the Kullback-Leibler (KL) Divergence between the prior state distribution and the posterior distribution derived from SCADA measurements.
- A significant divergence ( $\mathcal{K}_k \geq \tau_{KL}$ ) indicates a thermodynamic mismatch, flagging the presence of an attack that contradicts physical laws.
Time-Series Transformer (Temporal & Cross-Channel Classification):
- This layer processes sliding windows of multivariate data, including state estimates, sensor telemetry, and linear residuals.
- It utilizes Multi-Head Cross-Attention to analyze relationships between decoupled features (e.g., checking if a temperature change corresponds to the expected response in neutron flux and Xenon-135).
- The model classifies operational states into three categories: Normal (Safe), Uncoordinated Attack (uncoordinated offsets), and Stealth Attack (algebraically masked).

3. Key Contributions

Physics-Based Decoupling Proof: The paper establishes a mathematical proof that algebraic stealth attacks, while successful against linear estimators, inevitably violate the continuous non-linear dynamics of coupled reactor kinetics.
Hybrid Detection Architecture: The integration of a non-linear Particle Filter (for thermodynamic consistency) with a Time-Series Transformer (for temporal and cross-sensor correlation) creates a robust defense mechanism.
Comprehensive Simulation: The authors developed a simulation environment using MATLAB for the 5-state PRK model and PyTorch for the machine learning framework, generating over 629,000 time-indexed observations under various attack scenarios.

4. Experimental Results

The framework was evaluated on a dataset generated from a Pressurized Water Reactor (PWR) model under standard operational transients and two attack types: uncoordinated abrupt offsets and continuous algebraic drift.

Overall Performance: The proposed method achieved an overall accuracy of 99.90% and an F1-score of 0.9990 across 629,760 samples.
Stealth Attack Detection: The model successfully identified 186,580 stealth attack instances with a precision of 0.9987 and recall of 0.9980, demonstrating its ability to bypass the "blindness" of linear detectors.
Uncoordinated Attack Detection: Achieved perfect precision and recall (1.0000) for uncoordinated attacks.
Error Rates: The system produced minimal false positives (251) and false negatives (383) across the entire dataset.
Dynamic Response: In a specific scenario where an attacker drove reported temperatures down to -34°C, the Time-Series Transformer detected the physical divergence and triggered an alert before thermal thresholds were exceeded, whereas linear estimators failed to detect the anomaly.

5. Significance and Claims

The paper claims that its primary significance lies in advancing nuclear cybersecurity by moving beyond static, linear residual monitoring. By leveraging the inherent non-linear thermodynamic coupling of nuclear reactors, the proposed framework exposes a fundamental limitation of algebraic stealth attacks: they cannot simultaneously satisfy discrete linear constraints and continuous physical laws.

The authors assert that this dual-layer approach (Particle Filter + Transformer) effectively detects FDIA by identifying the "physical decoupling" that attackers cannot fake over time. The work suggests that such hybrid frameworks are essential for securing critical infrastructure where safety depends on the integrity of coupled multi-physics systems.