Dimensionally constrained adversarial attack and defense in wind power forecasting.

Area of Science:

• Artificial Intelligence
• Renewable Energy Systems
• Cybersecurity in AI

Background:

• Deep neural networks (DNNs) excel in wind power forecasting but are susceptible to adversarial attacks.
• Existing attacks prioritize effectiveness over stealth, leaving a gap in understanding subtle threats.

Purpose of the Study:

• To propose a novel adversarial attack (DC-MI-FGSM) focusing on stealthiness for wind power forecasting.
• To develop a robust defense mechanism (DAE) against these sophisticated adversarial attacks.

Main Methods:

• Developed a dimension-constrained momentum iterative fast gradient sign method (DC-MI-FGSM) for generating stealthy adversarial perturbations.
• Implemented a denoising autoencoder (DAE)-based preprocessing strategy to restore adversarial samples to their clean forms.
• Validated methods on the SDWPF dataset under white-box and black-box attack scenarios.

Main Results:

• DC-MI-FGSM demonstrated superior stealthiness (lower APP) and effectiveness (higher MAPE, RMSE, MAE degradation) compared to existing attacks.
• The DAE defense successfully mitigated adversarial perturbations, significantly reducing forecasting errors.
• DAE preprocessing outperformed adversarial training in robustness and usability for wind power forecasting.

Conclusions:

• The proposed DC-MI-FGSM presents a significant advancement in adversarial attack stealthiness for wind power forecasting.
• The DAE-based defense offers an effective and practical solution for enhancing the security of DNNs in this domain.
• This research highlights the critical need for stealth-aware adversarial attack and defense strategies in critical infrastructure like wind energy.