Work

SAAs: Vulnerability of Learning-Based Anomaly Detection to Stealthy Adversarial Attacks on Grid-Tied Photovoltaic Systems

Year: 2026

Type: article

Abstract: Learning-based anomaly detection is increasingly adopted in grid-tied photovoltaic (PV) systems. Advanced multi-variate detectors implicitly learn both temporal dynamics and inter-channel spatial correlations from tightly coupled sensor telemetry, forming a powerful defense against data manipulation. Conventional adversarial attacks (AAs), which treat the feature space as unconstrained, generate physically inconsistent “ghost data” that violate these correlations and are therefore readily detected. This paper systematically investigates whether such spatiotemporal-aware detectors remain vulnerable to carefully designed perturbations. We show what they do: the proposed stealthy adversarial attacks (SAAs), based on physics-constrained projected gradient descent, optimize perturbations within a root-feature subspace and reconstruct all dependent channels through a differentiable physical layer, such that every adversarial sample is confined to the physical manifold and preserves temporal and spatial consistency by construction. A manifold-geometric analysis explains why such on-manifold perturbations are indistin-guishable to any detector that flags anomalies by their departure from the learned benign manifold, and a controlled ablation isolates the spatial-algebraic and temporal-coherence mechanisms behind the evasion. Hardware-in-the-loop experiments against four detectors — including the spatiotemporal graph-based DST-GNN — show that SAAs achieve attack success rates exceeding 86%, drastically outperforming AAs. We further demonstrate that the attack remains effective under LCL-parameter mismatch and time-varying irradiance/temperature, generates each masked window within a single rolling window including 50 sample points on commodity hardware without a dedicated accelerator — although sustaining this at the full per-sample rate remains beyond a single device — and induces sustained, quantified over-voltage that the blinded detectors fail to alarm, revealing a critical vulnerability of residual/forecasting-based spatiotemporal anomaly detectors in grid-tied PV systems.

Source: IEEE Transactions on Smart Grid

DOI: 10.1109/tsg.2026.3722831

Authors: Haoming Liu, Mengxiang Liu, Sha Peng, Qiliang Jiang, Yitian Zhou, Ruilong Deng

Institutions: Zhejiang University, University of Bristol

Language: English

Topic: Smart Grid Security and Resilience

Subfield: Control and Systems Engineering

Field: Engineering

Domain: Physical Sciences

SDG: Affordable and clean energy

Open Access status: hybrid

Estimated APC paid: $2,645