Federated localization with fdia_graph.federated

Status. The results on this page are on data release v0.8.1, pinned in the run script, and predate the measurements-only feature rules (swing from observed frames, the Jacobian block against the previous frame's estimate, measured calibration) and the one-frame Aq and Al. They are not yet rerun on v0.8.3; the centralized numbers in ../localization/README.md are current.

The federated localization papers' protocol, run by the SDK on the v0.8.1 timelines (pinned in the run script) with the 14-dim per-bus vector plus the Jacobian block.

import fdia_graph as fg
from fdia_graph.federated import FedBusCNN

rel = dict(release="v0.8.1")  # the release these results are on
zs = dict(families=[0, 1, 2], **rel)  # benign + Aq + Ad
train = fg.load("ieee118", split="train", **zs)
val = fg.load("ieee118", split="val", **zs)
test = fg.load("ieee118", split="test", families=[0, 1, 2, 3, 4], **rel)  # adds As and Ar

loc = FedBusCNN(K=3, rounds=60, local_epochs=3, features="full14+jac").fit(train, val=val)
tab = loc.score_perbus(test, buses="attackable", fr_over="all")  # the paper's Table IV block

release= pins the built-in data only: a dataset registered locally under the same name loads in its place. fg.list_datasets()["ieee118"] reads "builtin" when these are the release's records, and the run script refuses to fit otherwise.

clients	K utilities from the spectral partition of the bus graph (random_state 42); each client reads only its own buses and its own flow meters for the power balance
training	FedAvg with uniform weights, 60 rounds of 3 local epochs, AdamW lr 5e-4, weight decay 0.01, batch 256, gradient clip 1.0, the final-round weights
features	the papers' 14-dim per-bus vector built per client, plus the 8-channel Jacobian block, computed once centrally and shared, the one feature a client does not build from its own meters
threshold	the papers' validation-best global tau on the 0.05 to 0.95 grid, from per-client confusion counts summed at the server
what crosses a boundary	per-channel moments once, the model weights every round, the per-bus confusion counts once, and the central Jacobian block

Results

These runs used the earlier Jacobian block, which took the change against the previous frame's true state. The block now reads observed data only (the previous frame's estimate), and these tables await a rerun on it.

Zero-shot: train and val hold benign, Aq and Ad; test adds As and Ar, never seen in training. F1, DR and FR are the paper's per-bus macro scores over the attackable buses, mean and standard deviation over seeds 123, 124 and 125. Every run is in docs/federated/results/runs/, the aggregates in the results store, experiment federated.localization (results/).

Table IV layout (FR over every test record, the paper's convention)

Model	F1 14	DR 14	FR 14	F1 118	DR 118	FR 118	F1 300	DR 300	FR 300
1D CNN, K = 1	0.878 ± 0.002	0.825 ± 0.009	0.0074 ± 0.0014	0.899 ± 0.002	0.848 ± 0.003	0.0005 ± 0.0000	0.847 ± 0.008	0.790 ± 0.018	0.0005 ± 0.0001
1D CNN, K = 2	0.860 ± 0.007	0.792 ± 0.015	0.0066 ± 0.0008	0.898 ± 0.001	0.847 ± 0.006	0.0005 ± 0.0001	0.861 ± 0.004	0.802 ± 0.011	0.0004 ± 0.0000
1D CNN, K = 3	0.850 ± 0.002	0.775 ± 0.006	0.0061 ± 0.0006	0.904 ± 0.001	0.853 ± 0.005	0.0004 ± 0.0001	0.861 ± 0.003	0.800 ± 0.006	0.0003 ± 0.0000
Per-bus MLP, K = 1	0.833 ± 0.016	0.750 ± 0.028	0.0064 ± 0.0009	0.885 ± 0.001	0.829 ± 0.002	0.0005 ± 0.0000	0.834 ± 0.004	0.791 ± 0.022	0.0007 ± 0.0002
Per-bus MLP, K = 2	0.828 ± 0.014	0.753 ± 0.027	0.0077 ± 0.0016	0.865 ± 0.004	0.802 ± 0.010	0.0006 ± 0.0001	0.821 ± 0.004	0.774 ± 0.008	0.0007 ± 0.0001
Per-bus MLP, K = 3	0.832 ± 0.010	0.770 ± 0.022	0.0102 ± 0.0017	0.861 ± 0.003	0.805 ± 0.008	0.0008 ± 0.0001	0.822 ± 0.001	0.776 ± 0.008	0.0006 ± 0.0001

FR over benign records only (the budget the localization guide reports)

Model	FR 14	FR 118	FR 300
1D CNN, K = 1	0.00000 ± 0.00001	0.00000 ± 0.00000	0.00000 ± 0.00000
1D CNN, K = 2	0.00000 ± 0.00001	0.00001 ± 0.00001	0.00001 ± 0.00000
1D CNN, K = 3	0.00014 ± 0.00014	0.00002 ± 0.00000	0.00001 ± 0.00000
Per-bus MLP, K = 1	0.00009 ± 0.00005	0.00002 ± 0.00000	0.00004 ± 0.00001
Per-bus MLP, K = 2	0.00010 ± 0.00006	0.00006 ± 0.00004	0.00014 ± 0.00003
Per-bus MLP, K = 3	0.00007 ± 0.00005	0.00007 ± 0.00001	0.00008 ± 0.00003

Per-family node F1 (that family plus benign, mean over seeds)

Model	Aq 14	Ad 14	As 14	Ar 14	Aq 118	Ad 118	As 118	Ar 118	Aq 300	Ad 300	As 300	Ar 300
1D CNN, K = 1	0.671	0.980	0.917	0.627	0.391	0.985	0.950	0.749	0.173	0.982	0.934	0.754
1D CNN, K = 2	0.656	0.965	0.900	0.585	0.384	0.982	0.942	0.757	0.219	0.977	0.925	0.769
1D CNN, K = 3	0.647	0.962	0.895	0.554	0.399	0.981	0.945	0.762	0.215	0.973	0.925	0.771
Per-bus MLP, K = 1	0.602	0.966	0.863	0.544	0.376	0.976	0.917	0.756	0.182	0.969	0.911	0.734
Per-bus MLP, K = 2	0.613	0.958	0.855	0.537	0.342	0.967	0.903	0.725	0.127	0.964	0.891	0.747
Per-bus MLP, K = 3	0.599	0.953	0.866	0.571	0.331	0.962	0.897	0.740	0.123	0.962	0.895	0.757

Per-bus F1 by attack family, row labels carrying each row's FR over every record.

IEEE 14	IEEE 118	IEEE 300

Three readings

reading	evidence	open case
federating costs little	from one client to three the CNN moves 0.878 to 0.850 on IEEE-14 and holds 0.899 to 0.904 on 118 and 0.847 to 0.861 on 300	the MLP loses a few points on 118 and 300 as K grows
the CNN is the arm to deploy	its macro F1 leads the per-bus MLP's at every K on every system, at a false-positive rate of the same small order (the Table IV layout)	the MLP is a third of the parameters (52k against 154k) for a few points of F1, and matches the CNN on Aq at K = 1 on 300
the stealthy re-solve falls with size	the CNN's Aq node F1 at K = 1 is 0.671, 0.391 and 0.173 on 14, 118 and 300 while Ad, As and Ar stay far higher (the per-family table)	a sustained local false state inside an episode, the same frontier as in ../localization/README.md; the v0.8.1 timelines hold Aq over multi-frame episodes, while the generator in this package makes every Aq episode one frame

The papers report CNN macro F1 of 0.963, 0.963 and 0.952 on the v0.4.1 record shards, and the SDK reproduces them within half a point on the v0.7.2 shards. On a timeline the 14-dim vector alone drops, because its temporal features compare each frame with the frame one minute before rather than with the benign scan before an attacked snapshot. The Jacobian block reads each frame's physics instead of its history, which recovers most of that drop on 118 and 300 and about half of it on 14; the centralized comparison with and without the block is in ../localization/README.md.

Regenerate

FG_SYSTEM=ieee14 python docs/federated/run_federated.py     # 18 runs, writes docs/federated/results/runs/ + a run of federated.localization
FG_SYSTEM=ieee118 python docs/federated/run_federated.py
FG_SYSTEM=ieee300 python docs/federated/run_federated.py
python docs/federated/make_report.py                        # figures + CSV sidecars from the store
python tools/results_docs.py --write                        # the README's results blocks

Saved runs are skipped, so a re-run only aggregates.