# FalsEye: proactive detection of false data injection attacks in smart grids using IceCube-optimised ensemble learning

## TL;DR

This paper introduces FalsEye, a new method to detect false data attacks in smart grids using optimized machine learning techniques.

## Contribution

The novel FalsEye framework combines ensemble learning with a physics-inspired optimization algorithm and adaptive oversampling for improved FDIA detection.

## Key findings

- The IO Voting Classifier achieved 99% accuracy and 95% F1-score in detecting false data injection attacks.
- The framework outperformed conventional methods in precision, recall, and F1-score metrics.
- Adaptive oversampling improved detection of minority FDIA instances in imbalanced datasets.

## Abstract

False Data Injection Attacks (FDIAs) represent a significant cybersecurity threat to smart grids (SGs), compromising both system stability and operational reliability. Conventional detection approaches frequently prove inadequate, largely due to challenges such as data imbalance and suboptimal model parameterisation. To overcome these limitations, this study proposes a proactive detection framework that integrates ensemble learning, adaptive oversampling, and a novel metaheuristic optimization algorithm, termed FalsEye. At the core of the proposed framework is a Voting Classifier ensemble, which strategically combines heterogeneous base learners, including ExtraTrees, CatBoost, and LightGBM. The performance of this ensemble is further enhanced through the IceCube Optimization (IO) algorithm, a physics-inspired metaheuristic technique employed to fine-tune the hyperparameters of the individual base models. Additionally, the framework incorporates adaptive oversampling using the Adaptive Synthetic method to effectively mitigate class imbalance within the dataset, thereby improving the detection rate of minority FDIA instances. Experimental results demonstrate that the IO Voting Classifier achieves superior F1-scores and exhibits a more balanced precision–recall trade-off compared to conventional ensemble approaches. The optimized framework attains an accuracy of 99%, with a precision of 92%, a recall of 98%, and an F1-score of 95%, marking a substantial improvement over traditional methods. These findings highlight the considerable potential of combining metaheuristic optimization with ensemble learning to develop robust and cyber-resilient SG infrastructures.

The online version contains supplementary material available at 10.1038/s41598-026-38723-0.