An AI-Driven Adversarial Attack and Defense Framework for Enhancing Cybersecurity Resilience in Smart Grid Systems

Smart grids rely on machine-learning detectors to flag cyber attacks in synchrophasor and control data. These detectors are accurate on clean data, yet their robustness to adversarial manipulation is rarely tested end to end. We build one framework that attacks and defends the same detectors under a single protocol. It runs on real synchrophasor measurements and on a verifiable false data injection simulation. On clean data a random forest reaches 92 percent accuracy on binary attack detection and 83 percent on 37-class detection. It reaches 100 percent on the injection task. White-box attacks drive the neural detector to chance-level accuracy or below. Black-box transfer from an independent surrogate drops the gradient-free random forest from 92 to 31 percent and from 83 to 6 percent. On the injection task it drops from 100 to 50 percent. Among four defenses, only adversarial training survives attacks adapted to it. It holds 99 percent robust accuracy on the injection task. On the binary synchrophasor task its 71 percent only matches the attack base rate. Distillation, ensembles, and feature squeezing fail under adapted attacks. A leakage-robust split lowers synchrophasor accuracy to the base rate, which shows that reported accuracy on this dataset is partly an artifact of correlated samples.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-19
DOI
https://doi.org/10.21203/rs.3.rs-11518390/v1
Primary Topic
Smart Grid Security and Resilience
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

An AI-Driven Adversarial Attack and Defense Framework for Enhancing Cybersecurity Resilience in Smart Grid Systems

Jing Zeng, Pingzhou Li, Shuo Li
Research Square
Smart Grid Security and Resilience
preprint

An AI-Driven Adversarial Attack and Defense Framework for Enhancing Cybersecurity Resilience in Smart Grid Systems

Jing Zeng, Pingzhou Li, Shuo Li
preprint en

Abstract

Smart grids rely on machine-learning detectors to flag cyber attacks in synchrophasor and control data. These detectors are accurate on clean data, yet their robustness to adversarial manipulation is rarely tested end to end. We build one framework that attacks and defends the same detectors under a single protocol. It runs on real synchrophasor measurements and on a verifiable false data injection simulation. On clean data a random forest reaches 92 percent accuracy on binary attack detection and 83 percent on 37-class detection. It reaches 100 percent on the injection task. White-box attacks drive the neural detector to chance-level accuracy or below. Black-box transfer from an independent surrogate drops the gradient-free random forest from 92 to 31 percent and from 83 to 6 percent. On the injection task it drops from 100 to 50 percent. Among four defenses, only adversarial training survives attacks adapted to it. It holds 99 percent robust accuracy on the injection task. On the binary synchrophasor task its 71 percent only matches the attack base rate. Distillation, ensembles, and feature squeezing fail under adapted attacks. A leakage-robust split lowers synchrophasor accuracy to the base rate, which shows that reported accuracy on this dataset is partly an artifact of correlated samples.

Research Square
State Grid Corporation of China (China) (CN)
Smart Grid Security and Resilience
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon