Configuration Reference¶

Model training and evaluation are configured via YAML files. Config files for the Container Quickstart workflow live in configs_container_qs/; configs used for the paper's experiments live in configs/.

Example¶

# configs/ps_attack/exp1_ps_attack_ec1e1.yaml

data_file_name: "./data/ps_attack_with_timestamps.csv"
result_root:    "./results/for_paper/ps_attack/multirun/ec1e1_gc1em2_gr1em2/"

state_dict:   ["R1-PA1:VH", "R1-PM1:V", "R1-PA2:VH", "R1-PM2:V"]
measure_dict: ["R1-PA4:IH", "R1-PM4:I", "R1-PA5:IH", "R1-PM5:I"]

LSTMhiddensize: 20
LSTMlayers:      1
FFhiddensize:    5
num_epoch:     650
batch_size:      1

la_mse:      1
la_nll_r:    0
la_nll_h:    1
la_CPCT_cov: 1
la_K:        0
lr:          1e-3

lookback_bptt: 128

init_time:          "2020-01-01 00:00:01"
start_time:         "2020-01-01 00:00:01"
attack_start_time:  "2020-01-01 00:02:53"
attack_id:          "ps1"
attack_duration:    426
duration:           600

training_data_fraction: 1.0
test_window:     150
forecast_window: 155
type:            'any'

Fields¶

Data¶

Key	Type	Description
`data_file_name`	str	Path to the input CSV dataset
`result_root`	str	Directory where results and model weights are written
`state_dict`	list[str]	Column names in the CSV that are treated as EKF state variables
`measure_dict`	list[str]	Column names in the CSV that are treated as observations (measurements)

Model architecture¶

Key	Type	Description
`LSTMhiddensize`	int	Hidden size of the LSTM state transition sub-model
`LSTMlayers`	int	Number of stacked LSTM layers
`FFhiddensize`	int	Hidden size of the feed-forward observation decoder

Training¶

Key	Type	Description
`num_epoch`	int	Number of training epochs
`batch_size`	int	Batch size (typically 1 for sequential BPTT)
`lr`	float	Adam learning rate
`lookback_bptt`	int	BPTT window length — the dataset is sliced into chunks of this length and shuffled across epochs
`training_data_fraction`	float	Fraction of the training split to actually use (1.0 = all)

Loss weights¶

Key	Type	Description
`la_mse`	float	Weight on the state MSE loss term
`la_nll_r`	float	Weight on the observation NLL loss term
`la_nll_h`	float	Weight on the state NLL loss term
`la_CPCT_cov`	float	Weight on the innovation covariance NLL loss term
`la_K`	float	Weight on the Kalman gain consistency loss term

Temporal windowing¶

Key	Type	Description
`init_time`	str	Timestamp of the first row in the dataset (ISO 8601)
`start_time`	str	Timestamp to begin evaluation from
`attack_start_time`	str	Timestamp of the first injected attack sample
`attack_id`	str	Identifier string for the attack scenario
`attack_duration`	int	Duration of the attack in timesteps
`duration`	int	Total evaluation duration in timesteps
`test_window`	int	Number of timesteps in the test split
`forecast_window`	int	Number of timesteps used for forecast evaluation
`type`	str	Evaluation type tag (e.g. `'any'`)

Environment variables¶

The Trainer class also reads the following environment variables at runtime, which override config-file values where applicable:

Variable	Default	Description
`D_VAL`	`120`	DVAL interval length (EKF steps per proof)
`HYP_TEST_INT`	`128`	Hypothesis test interval in timesteps
`GEN_PREDICTION_PROOFS`	`1`	Set to `0` to skip prediction proof generation
`GEN_HYPOTHESIS_PROOFS`	`1`	Set to `0` to skip hypothesis proof generation
`PRELOADED_PROOF_ROOT`	—	Path where VK/PK artefacts are stored and read from. Always set by `zkstarctl.sh` to `results/proofs/<dataset>/<config>/`. After `utility-setup` completes, a zero-byte tolerance marker file (`tol.svd=<val>.kcr=<val>`) is written here as a permanent record of the circuit parameters used.
`MAX_SVD_TOL`	`0.1`	Absolute tolerance (`atol`) for `SVDReconstructionCheck` inside the hypothesis circuit. Baked into the circuit at `utility-setup` time; has no effect if passed to `utility-gen-witness` or `utility-prover`.
`MAX_KCR_TOL`	`0.1`	Absolute tolerance (`atol`) for `KalmanReconstructionCheck` inside the prediction circuit. Baked into the circuit at `utility-setup` time; has no effect if passed to `utility-gen-witness` or `utility-prover`.
`ALPHA`	`0.005`	Significance level α for the chi-squared test
`NUM_THREADS`	—	PyTorch CPU thread count
`TRAIN_RATIO`	`0.8`	Fraction of data used for training
`WRITE_WEIGHT_FILE`	UUID-based	Output path for trained model weights
`VECTORIZED`	`1`	Set to `0` to disable vectorized covariance updates
`DIAG_COV`	—	Set to use diagonal-only covariance approximation