fdia-graph

Stealthy FDIA localization datasets for power grids, PyTorch-ready in one line. Eight IEEE systems (14 / 30 / 57 / 89 / 118 / 145 / 200 / 300 buses), 72,000 records each.

import fdia_graph as fg

ds = fg.load("ieee118", split="train")     # auto-downloads + caches
loader = ds.loader(batch_size=64)
for batch in loader:
    batch["node_x"], batch["edge_x"], batch["edge_index"], batch["y"], batch["family"]

New here?

Install

pip install fdia-graph              # loader
pip install "fdia-graph[torch]"     # + PyTorch DataLoader
pip install "fdia-graph[pyg]"       # + torch_geometric
pip install "fdia-graph[se]"        # + state estimation / residual localization (torch + pandapower)
pip install "fdia-graph[generate]"  # + pandapower, to generate custom data

Data is pinned per SDK version and cached in ~/.cache/fdia_graph. Pin a data version with fg.load(..., release="v0.7.2"); pip install --upgrade fdia-graph moves it forward.

Load

fg.load("ieee300", split="train")                              # 60/20/20 chronological split
fg.load("ieee118", split="test", families=["Aq","At","Al"])    # family subset
fg.load("ieee118", units="pu")                                 # per-unit + radians (default is physical)

• Whole split at once: ds.to_numpy() / .to_torch() / .to_pandas().
• Custom data: fg.generate(system, name, per_family=..., attack_intensity=..., ...) then fg.load(name).
• Continuous timeline for LSTM/TGN: fg.load_stream(system).

Details for all three in docs/reference/EXAMPLES.md.

State estimation

from fdia_graph.se import WLS, SubspacePrior   # pip install "fdia-graph[se]"

train, test = fg.load("ieee118", split="train"), fg.load("ieee118", split="test")
est = SubspacePrior(rank_frac=0.5, reweight="huber", c=2.5).fit(train)
xhat = est.estimate(test)          # [n, 2N-1] = [theta rad (non-slack) | V pu (all buses)]
print(est.score(test))             # per-family angle/voltage MAE vs the clean truth

WLS, AdaptiveWeighting, ResidualRemoval, SubspacePrior share one solver and differ only in state space and weights. Results and a walkthrough: docs/se/.

Localization

from fdia_graph.localization import SwingThreshold   # numpy only

loc = SwingThreshold(fa_target=0.01).fit(train)     # per-bus thresholds from benign records only
flag = loc.localize(test)                           # [n, N] bool: which buses are called attacked
print(loc.score(test))                              # per-family node-F1, strict accuracy, DR next to FA

SwingThreshold, DeltaThreshold, ResidualLocalizer share one calibration and metric protocol and differ only in the per-bus score. BusCNN and BusMLP are the papers' learned localizers on the 14-dim per-bus feature vector (needs [torch]):

from fdia_graph.localization import BusCNN

zs = dict(families=[0, 1, 2])                        # papers' zero-shot protocol: As/Ar unseen in train
cnn = BusCNN().fit(fg.load("ieee118", split="train", **zs), val=fg.load("ieee118", split="val", **zs))
print(cnn.score(fg.load("ieee118", split="test", families=[0, 1, 2, 3, 4]))["all"]["macro_f1"])

Results: docs/localization/.

Data

Each record is a sparse measurement graph with N buses (nodes) and E branches (edges). Read a shape as "values per item": [N,4] = 4 numbers per bus, [E,8] = an 8-dim vector per branch.

In PyG format (format="pyg" and fg.pyg_stream) the flows are Data.edge_attr (alias Data.edge_x), the mask is edge_mask, and the static line physics are Data.edge_phys. Full reference: docs/reference/DATA_DICTIONARY.md.

Attacks

Three stealthy families that evade classical bad-data detection (BDD), plus three detectable ones as a contrast set.

Bad-data statistic J relative to the alarm threshold, per family (Aq is labeled A_o here). Green families are indistinguishable from benign; red ones trip the alarm.

• Every per-bus change stays in a plausibility band: 2% noise floor to 20% cap.
• Meter error follows an accuracy-class model (per-meter bias plus per-scan jitter).
• Report per-family node-F1 with the false-alarm rate, not accuracy: clean buses dominate.
• A lightweight per-bus MLP reaches ~0.92 localization macro-F1 (see the examples page).

Citation

Cite the attack- and measurement-model sources:

• Yuan, Li & Ren, Modeling load redistribution attacks in power systems, IEEE T-SG 2(2), 2011. (LRA)
• Haghshenas, Hasnat & Naeini, A Temporal GNN for Cyber Attack Detection and Localization in Smart Grids, IEEE ISGT 2023. (ramp)
• Zaman & Lin, PING: Physics-Informed GNNs to Generalize FDIA Localization, NAPS 2025. (measurement model)
• Asprou, Kyriakides & Albu, Variable Weights in a WLS State Estimator, IEEE T-IM 63, 2014. (meter noise)
• Boyaci et al., Joint Detection and Localization of Stealth FDIA, IEEE T-SG, 2022. (protocol)

License

Data under CC BY 4.0, code under MIT (see LICENSE). Synthetic, from public IEEE cases. Not for operational decisions.