Enhancing Smart Grid Cybersecurity: Multiobjective Optimization of Informer-Based FDIA Detection With Adversarial Defense Mechanisms

Abstract

In the era of smart grids, the use of digital technologies has introduced unprecedented cybersecurity challenges, particularly in safeguarding against false data injection attacks (FDIAs) that may potentially manipulate power system measurements and disrupt grid operation. This article proposes an informer-based FDIA detection framework with multiobjective optimization, a nondominated sorting genetic algorithm II (NSGA-II), and adversarial defense techniques to maximize the attack detection rate (ADR), minimize the false positive rate (FPR), and reduce computational overhead. Leveraging ProbSparse self-attention, the informer efficiently captures long-range temporal dependencies with low complexity, making it well-suited for real-time anomaly detection in power-system time series. Experimental evaluation on IEEE 5-bus, 14-bus, and 118-bus systems demonstrates the robustness and scalability of the proposed approach. Feature selection optimization, evaluated relative to the same Informer detector trained without feature selection, increased the ADR by 12.5 percentage points and reduced the FPR by 9 percentage points, confirming that redundant features should be omitted. Specifically, Bayesian hyperparameter optimization reduces informer training time from 188.7 to 145.2 s (22.0%), while subsequent NSGA-II-based multiobjective optimization further improves ADR from 87.3% to 95.1% under identical training conditions. The proposed defense mechanism is robust against fast gradient sign method (FGSM), basic iterative method (BIM), and momentum iterative method (MIM) attacks, and it outperforms a conventional autoencoder-based detector. Moreover, the proposed detector consistently outperforms state-of-the-art transformer time-series baselines (TimesNet, PatchTST, and FEDformer) under the same experimental protocol. Real-time feasibility experiments further show that the informer achieves an inference latency of under 6 ms per sample and a maximum throughput of 2100 samples per second, supporting deployment in large-scale smart grid monitoring.