Multi-task transformer with adaptive attention and uncertainty-aware recovery for FDI attacks defense in smart grids

Abstract

False Data Injection (FDI) attacks pose significant threats to the stability and security of modern power systems by covertly altering measurement data. This study focuses on stealthy FDI attacks, where an attacker subtly manipulates measurements within a localized region, leveraging limited network information to overload specific transmission lines. To address this challenge, a deep learning framework based on Multi-Task Transformer with Adaptive Attention and Uncertainty-Aware Recovery (MTT-AUR) is proposed. MTT-AUR performs simultaneous attack detection, region and bus-level localization, and measurement recovery solely based on snapshot measurements. MTT-AUR integrates adaptive multi-scale attention and an uncertainty-aware recovery module with a gating mechanism, which enables precise reconstruction of compromised measurements. A multi-task loss function based on uncertainty dynamically balances the four tasks, optimized through variable coefficients determined by each task's uncertainty. Simulations on the IEEE 39-bus system were conducted to evaluate and compare the performance of MTT-AUR against existing deep learning and machine learning methods. Additional experiments on the IEEE 118-bus system demonstrate the scalability and robustness of the model across all four tasks. © 2026 Elsevier B.V.