🤖 machine learning

Attention-Enhanced Graph Filtering for False Data Injection Attack Detection and Localization

This paper proposes a novel framework that integrates ARMA-based graph filters with an Encoder-Only Transformer to effectively detect and localize False Data Injection Attacks in power systems by simultaneously capturing local structural dependencies and long-range global contextual relationships, as validated on NYISO data across IEEE 14- and 300-bus systems.

Original authors: Ruslan Abdulin, Mohammad Rasoul Narimani

Published 2026-01-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ruslan Abdulin, Mohammad Rasoul Narimani

Original paper licensed under CC BY 4.0 (http://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

Imagine a massive, interconnected city where electricity flows through a complex web of power lines, much like water flowing through a city's plumbing system. To keep the lights on and the system safe, the city uses thousands of sensors (like smart meters) to constantly check the pressure and flow of electricity. These sensors send their reports to a central "brain" (the control center) that tries to figure out exactly what the system is doing at any given moment.

The Problem: The Sneaky Prankster
The paper describes a specific type of cyberattack called a "False Data Injection Attack" (FDIA). Imagine a prankster who doesn't break into the control center but instead sneaks into the sensor reports. They tweak the numbers just enough to look real, but they lie about the actual state of the system.

Traditional security guards (old detection methods) look for obvious errors, like a number that is way too high or low. But this prankster is smart; they change the numbers in a way that still follows the laws of physics. To the old guards, the data looks perfect, even though it's a lie. This is dangerous because the control center might make decisions based on fake information, potentially causing blackouts or equipment damage.

The Old Solutions: The Flawed Detectives
Researchers have tried using Artificial Intelligence (AI) to catch these liars.

The "Local" Detective: Some AI models look at a sensor and its immediate neighbors. They are good at spotting local trouble but can get confused if the lie is part of a bigger, coordinated pattern happening far away.
The "Global" Detective: Other models try to look at the whole city at once. However, if they try to look at everything too deeply, they sometimes get "blurry" and lose track of exactly where the problem started. They might know something is wrong but can't point to the specific broken pipe.

The New Solution: The "ACEOT" Team
The authors of this paper propose a new, hybrid detective team called ACEOT (ARMAConv Encoder-Only Transformer). Think of it as combining two specialized skills into one super-agent:

The "Map Reader" (ARMAConv): This part of the team is an expert at reading the city's map. It understands the physical connections between power lines and sensors. It uses a special technique (called ARMA filtering) to smooth out the noise and focus on the immediate neighborhood. It's great at saying, "Hey, this specific block looks suspicious compared to the houses right next to it."
The "Big Picture" Detective (Transformer): This part uses a mechanism called "Self-Attention." Imagine a detective who can instantly zoom out and see how a problem in the north district might be connected to a weird event in the south district, even if they are miles apart. This helps catch coordinated attacks that try to hide by spreading the lie across the whole city.

How They Work Together
The magic of ACEOT is how these two parts talk to each other:

First, the Map Reader cleans up the data and makes sure every sensor keeps its unique identity (so the AI doesn't get confused about which sensor is which).
Then, the Big Picture Detective looks at the cleaned-up data and asks, "Does this pattern make sense when I look at the entire grid?"
Finally, the system does two things at once:
1. Detection: It raises a red flag if any part of the city is under attack.
2. Localization: It points a finger at the exact sensors that are lying, so the operators know exactly where to send the repair crew.

The Results: A Smarter Guard
The researchers tested this new system on two different-sized "cities" (simulated power grids with 14 and 300 nodes) using real-world data from New York's power grid.

Accuracy: The new system was better at finding the liars than the old AI models.
Precision: It was very good at not crying "wolf" when everything was actually fine (low false alarms).
Scalability: The bigger the city (the more power lines), the better the new system performed compared to the old ones. This is because the "Big Picture Detective" gets more useful when there are more distant connections to analyze.

In a Nutshell
This paper introduces a smarter way to catch cyber-attacks on power grids. Instead of just looking at neighbors or just looking at the whole map separately, the new system combines both. It understands the local neighborhood and the global city layout simultaneously, allowing it to spot sneaky lies and pinpoint exactly where they are happening, keeping the power grid safer and more reliable.

Technical Summary: Attention-Enhanced Graph Filtering for False Data Injection Attack Detection and Localization

Problem Statement
The increasing integration of Internet-of-Things (IoT) devices in modern power grids has expanded the cyberattack surface, making False Data Injection Attacks (FDIAs) a critical threat. FDIAs are designed to bypass traditional model-based detection schemes, such as Bad Data Detection (BDD) based on the largest normalized residual test, by crafting measurement vectors that satisfy the physical power flow equations. While existing data-driven approaches utilizing Graph Neural Networks (GNNs) have shown promise, they face specific limitations:

Structural Limitations: Conventional GNNs (e.g., GCNs) often rely on first-order spectral approximations, leading to high computational costs and limited receptive fields. Higher-order variants like Chebyshev Graph Convolutional Networks (CGCNs) suffer from over-smoothing and noise amplification, which obscures the distinction between attacked and healthy nodes, hindering precise localization.
Rigid Dependencies: Standard deep learning models often fail to explicitly account for the underlying network topology or rely on rigid, input-invariant dependency patterns, limiting their adaptability to heterogeneous and spatially varying attack scenarios.
Transformer Limitations: While Transformer architectures can model long-range dependencies, their standalone application in power systems is computationally demanding and sensitive to noisy inputs. Furthermore, naively applying deep Transformer stacks can struggle to model localized grid dynamics effectively.

The core challenge addressed is the need for a framework that jointly detects the presence of an FDIA at the system level and localizes compromised buses at the node level, balancing robust local feature extraction with global contextual modeling.

Methodology: The ACEOT Framework
The authors propose ACEOT (ARMAConv Encoder-Only Transformer), a hybrid framework that integrates topology-aware graph filtering with self-attention mechanisms. The architecture consists of three primary components:

Learnable Positional Encoding (PE):
Unlike standard sinusoidal encodings designed for linear sequences, ACEOT employs a learnable embedding layer to assign unique identity vectors to each bus. This is critical because neighboring buses in a power grid may exhibit similar electrical characteristics; without explicit node identity, spectral filtering can lead to feature homogenization, masking subtle anomalies. These embeddings are added to node features before graph convolution to preserve discriminability.
ARMAConv Graph Filtering:
To extract robust, topology-aware local features, the framework utilizes Auto-Regressive Moving Average (ARMA) graph convolutional filters. Unlike polynomial-based filters (e.g., Chebyshev), ARMAConv employs learnable spectral filters that provide depth tolerance and enhanced noise robustness. The mechanism uses a recursive message-passing rule with parallel stacks to approximate the inverse of the graph shift operator iteratively. This approach mitigates over-smoothing and noise amplification, preserving node-level identity essential for localization.
Encoder-Only Transformer:
To capture global, input-dependent interactions without the computational overhead of a decoder, ACEOT employs an Encoder-Only Transformer. This component utilizes multi-head self-attention to model long-range dependencies among buses that may not be captured by localized spectral filtering alone. The self-attention mechanism allows the model to adaptively weight pairwise dependencies, identifying coordinated FDIA patterns spanning multiple grid regions.

Integration and Workflow
The processing pipeline begins with raw power injection measurements ( $P, Q$ ) and node indices. The indices are mapped to learnable positional embeddings, which are summed with the projected node features. This combined representation is passed through ARMAConv layers for topology-aware denoising and local aggregation. The output is then fed into the Encoder-Only Transformer to capture global context. Finally, a linear classification layer with sigmoid activation produces node-level attack probabilities. A graph-level detection output is derived by taking the maximum of these node probabilities.

Key Contributions

Hybrid Architecture: The proposal of a joint detection and localization framework combining ARMAConv for robust local feature extraction and an Encoder-Only Transformer for adaptive global dependency modeling.
Positional Encoding Integration: The integration of learnable positional encoding with ARMA-based filtering to provide persistent node identity, mitigating over-smoothing and noise amplification caused by malicious data injections.
Adaptive Localization: An attention-enhanced mechanism that enables input-dependent modeling of inter-node dependencies, improving localization accuracy across diverse and spatially distributed FDIA scenarios.
Comprehensive Validation: Extensive evaluation on IEEE-14 and IEEE-300 bus systems using real-world load data from the New York Independent System Operator (NYISO).

Experimental Results
The model was evaluated on synthetic datasets generated from NYISO load profiles (July 2021) for IEEE-14 and IEEE-300 bus systems, simulating four FDIA types (constrained optimization, distribution-based, data scaling, and data replay).

Detection Performance: In the IEEE-300 system, ACEOT achieved the highest F1 score (92.60%) among all benchmarks, including MLP, CNN, LSTM, ChebConv, and standalone ARMAConv. In the IEEE-14 system, it performed competitively, achieving an F1 score of 92.42%, slightly below ChebConv (92.46%) but with a significantly lower False Alarm Rate (1.28% vs. 1.32%).
Localization Performance:
- Sample-Wise: ACEOT significantly outperformed other models in the IEEE-300 system, achieving an 11.51% higher proportion of samples with F1 scores $\ge$ 95% compared to the next-best model (ARMAConv). This indicates the self-attention mechanism effectively captures long-range dependencies in large-scale networks.
- Node-Wise: ACEOT achieved the highest node-wise localization performance in both test systems, correctly classifying 49.09% of buses in the IEEE-14 system and showing superior trends in the IEEE-300 system.
Efficiency: While ACEOT is more complex than non-graph baselines, its inference time (approx. 20.96 ms for IEEE-300) is comparable to standalone ARMAConv (20.49 ms) and significantly faster than ChebConv (27.76 ms), demonstrating that the integration of the Transformer encoder does not introduce a prohibitive computational burden.

Significance and Claims
The paper claims that ACEOT offers a practical and effective solution for online FDIA detection and localization by leveraging both the state and topology of the power grid. The authors emphasize that the advantages of attention-based modeling become more pronounced as network size increases, highlighting the scalability of the approach. By combining structure-aware filtering with adaptive global attention, the framework alleviates over-smoothing and enhances robustness to noise, addressing the limitations of existing GNNs and deep learning models. The results suggest that ACEOT can reliably exploit global contextual information and long-range correlations, making it suitable for securing large-scale power systems against stealthy cyberattacks.

Technical Summary: Attention-Enhanced Graph Filtering for False Data Injection Attack Detection and Localization

More like this