← Lab /Security & privacy

Defense-aware XAI better identifies vulnerable OPF features than standard explainability

Security & privacy2026-10-28T21:31:05Z Measurement · Prototype · arXiv Threat model

The claim

The proposed framework more accurately identifies security-relevant vulnerabilities in OPF surrogate models than conventional explainability approaches

S. M. Monzurul Hoque Chowdhury, Qinghua Li, Jianhui Wang · arXiv Paper PDF

Published as: “A Defense-Aware Explainable AI Framework for Feature-Level Vulnerability Analysis in Optimal Power Flow”

So what

Enables operators to prioritize monitoring and hardening of specific load measurements that are both impactful on dispatch and hard to detect, improving cyber-resilience of OPF workflows.

The five-minute version

The proposed defense-aware XAI framework identifies vulnerable OPF input features far more accurately than standard OPF-SHAP importance, achieving 80% top-10% overlap with reference rankings on the IEEE 30-bus case and over 85% top-20% overlap on the IEEE 57-bus case. Evidence comes from comparisons against six anomaly-detection models on three IEEE test systems using a synthetic AC-OPF dataset. The paper’s results for the largest 118-bus system are truncated, so the claim of consistent scalability lacks full support.

What they found 3

Proposed framework achieves 80% overlap in top-10% vulnerable features on IEEE 30-bus system.
“the proposed framework consistently identifies vulnerable features across all six detection models, achieving an 80% overlap in the top-10% feature set.”
Baseline OPF-SHAP overlap varies between 20% and 60% on IEEE 30-bus system.
“the baseline OPF-SHAP ranking shows moderate but inconsistent agreement with the reference vulnerability rankings. As shown in Table I, the overlap between OPF-SHAP and the reference rankings ranges from 20% to 60% across different detection models.”
Baseline OPF-SHAP overlap is below 35% (top-10%) and 42% (top-20%) on IEEE 57-bus system.
“the OPF-SHAP ranking exhibits relatively low agreement with the reference vulnerability rankings, achieving less than 35% overlap in the top-10% feature set and below 42% overlap in the top-20% feature set across all detection models.”

What undercuts it the paper’s own results

The baseline OPF-SHAP method still attains up to 60% overlap in some cases, indicating that not all high-importance features are non-vulnerable.
Results for the IEEE 118-bus system are incomplete in the provided text, so the claim of consistent improvement at large scale is not fully substantiated.

What’s new here 4

Jointly models operational impact (OPF-SHAP) and detectability (autoencoder-SHAP) to produce a vulnerability score.
Introduces a ratio-based vulnerability metric Vi = φOPF i / φAE i.
Evaluates rankings against multiple independent reference rankings generated from six different detection models.
Demonstrates improved feature-ranking overlap across small, medium, and large IEEE test systems.

Threat model stated by the source

AdversaryAttacker capable of perturbing load measurement inputs

CapabilityCan modify individual input features within bounded physical limits

AssumesAttacker does not know the specific anomaly-detection model or its parameters; Perturbations must remain undetected by the deployed detector

Artifacts & methods 9

OPF surrogate neural networksystem
SHAP (Shapley value) explainabilityprimitive
Vanilla autoencodersystem
Contrastive Denoising Autoencodersystem
Sparse Autoencodersystem
Adversarial Autoencodersystem
Linear One-Class SVMsystem
Isolation Forestsystem
Synthetic AC-OPF dataset (National Laboratory of the Rockies)dataset

Topics

optimal power flowSHAP explainabilityautoencoder anomaly detectionfeature-level vulnerabilitydefense-aware machine learninginput-to-input relationshipscyber-physical attack detection

← Previous KeyDrift malware steals SD-WAN control-plane keys via in-process injection