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), one continuous timeline of 72,000 frames each: a record table when shuffled, a time series when not.

import fdia_graph as fg

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

ts = fg.load("ieee118", split="test", order="time")   # the same frames in time order
Xw, yw = ts.windows(W=24, stride=12)                  # [n, 24, N, 4] windows for an LSTM / TGN

Install

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

Load

One file per system, one loader. order decides what the file is to you:

fg.load("ieee300", split="train")                              # 60/20/20 chronological split, episodes never cut
fg.load("ieee118", split="test", families=["Aq", "At", "Al"])  # family subset
fg.load("ieee118", units="pu")                                 # per-unit + radians (default: physical)
fg.load("ieee118", order="random", seed=0)                     # the record table: a fixed permutation
fg.load("ieee118", split="test", order="time")                 # the time series: windows, episodes

State estimation

from fdia_graph.se import 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

One solver, six estimators that each change one thing: WLS, AdaptiveWeighting, ResidualRemoval, SubspacePrior, JacobianWeighting, GatedPrior. Results: docs/se/.

Localization

from fdia_graph.localization import SwingThreshold, BusCNN     # numpy only / [torch]

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

zs = dict(families=[0, 1, 2])                       # the papers' zero-shot protocol
cnn = BusCNN().fit(fg.load("ieee118", split="train", **zs), val=fg.load("ieee118", split="val", **zs))

One calibration, five localizers: SwingThreshold, DeltaThreshold, ResidualLocalizer, BusMLP, BusCNN. Results: docs/localization/.

Data

Each record is a sparse measurement graph with N buses (nodes) and E branches (edges). A shape reads "values per item": [N,4] is 4 numbers per bus.

Full reference: docs/reference/DATA_DICTIONARY.md.

Attacks

| Ad / As / Ar | meter bias / scaling / replay | caught | same band |

Every stealthy family is a local false state (Wu et al. 2026): the attacker solves the power flow of a subnetwork around the attack with the boundary voltages held true, writes only that subnetwork's meters, and the measurement vector stays consistent with an AC state, so the residual test sees noise. The meters written are the tamper masks in the file's attack/ group.

Bad-data statistic relative to the alarm threshold, per family. Green families are indistinguishable from benign; red ones trip the alarm. Meter error follows an accuracy-class model (per-meter bias plus per-scan jitter). Report per-family node-F1 next to the false-alarm rate, never accuracy.

Citation

License

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