README.md

---
license: cc-by-4.0
task_categories:
  - tabular-classification
  - graph-ml
tags:
  - intrusion-detection
  - CAN-bus
  - graph-neural-networks
  - knowledge-distillation
pretty_name: GraphIDS Paper Data
---

# GraphIDS — Paper Data

Evaluation artifacts for "Adaptive Fusion of Graph-Based Ensembles for Automotive Intrusion Detection".

Consumed by [kd-gat-paper](https://github.com/frenken-lab/kd-gat-paper) via `data/pull_data.py`.

## Schema Contract

**If you change column names or file structure, `pull_data.py` will fail.**
The input schema is enforced in `pull_data.py:INPUT_SCHEMA`.

### metrics.parquet

Per-model evaluation metrics across all runs.

| Column | Type | Description |
|---|---|---|
| `run_id` | str | Run identifier, e.g. `hcrl_sa/eval_large_evaluation` |
| `model` | str | Model name: `gat`, `vgae`, `fusion` |
| `accuracy` | float | Classification accuracy |
| `precision` | float | Precision |
| `recall` | float | Recall (sensitivity) |
| `f1` | float | F1 score |
| `specificity` | float | Specificity (TNR) |
| `balanced_accuracy` | float | Balanced accuracy |
| `mcc` | float | Matthews correlation coefficient |
| `fpr` | float | False positive rate |
| `fnr` | float | False negative rate |
| `auc` | float | Area under ROC curve |
| `n_samples` | float | Number of evaluation samples |
| `dataset` | str | Dataset name: `hcrl_sa`, `hcrl_ch`, `set_01`–`set_04` |

### embeddings.parquet

2D UMAP projections of graph embeddings per model.

| Column | Type | Description |
|---|---|---|
| `run_id` | str | Run identifier |
| `model` | str | Model that produced the embedding: `gat`, `vgae` |
| `x` | float | UMAP dimension 1 |
| `y` | float | UMAP dimension 2 |
| `label` | int | Ground truth: 0 = normal, 1 = attack |

### cka_similarity.parquet

CKA similarity between teacher and student layers (KD runs only).

| Column | Type | Description |
|---|---|---|
| `run_id` | str | Run identifier (only `*_kd` runs) |
| `dataset` | str | Dataset name |
| `teacher_layer` | str | Teacher layer name, e.g. `Teacher L1` |
| `student_layer` | str | Student layer name, e.g. `Student L1` |
| `similarity` | float | CKA similarity score (0–1) |

### dqn_policy.parquet

DQN fusion weight (alpha) per evaluated graph.

| Column | Type | Description |
|---|---|---|
| `run_id` | str | Run identifier |
| `dataset` | str | Dataset name |
| `scale` | str | Model scale: `large`, `small` |
| `has_kd` | int | Whether KD was used: 0 or 1 |
| `action_idx` | int | Graph index within the evaluation set |
| `alpha` | float | Fusion weight (0 = full VGAE, 1 = full GAT) |

**Note:** Lacks per-graph `label` and `attack_type`. The paper figure needs these fields joined from evaluation results. This is a known gap — see `pull_data.py` skip message.

### recon_errors.parquet

VGAE reconstruction error per evaluated graph.

| Column | Type | Description |
|---|---|---|
| `run_id` | str | Run identifier |
| `error` | float | Scalar reconstruction error |
| `label` | int | Ground truth: 0 = normal, 1 = attack |

**Note:** Single scalar error — no per-component decomposition (Node Recon, CAN ID, Neighbor, KL). The paper figure needs the component breakdown. This is a known gap.

### attention_weights.parquet

Mean GAT attention weights aggregated per head.

| Column | Type | Description |
|---|---|---|
| `run_id` | str | Run identifier |
| `sample_idx` | int | Graph sample index |
| `label` | int | Ground truth: 0 = normal, 1 = attack |
| `layer` | int | GAT layer index |
| `head` | int | Attention head index |
| `mean_alpha` | float | Mean attention weight for this head |

**Note:** Aggregated per-head, not per-edge. The paper figure needs per-edge attention weights with node positions. This is a known gap.

### graph_samples.json

Raw CAN bus graph instances with node/edge features.

Top-level keys: `schema_version`, `exported_at`, `data`, `feature_names`.

Each item in `data`:
- `dataset`: str — dataset name
- `label`: int — 0/1
- `attack_type`: int — attack type code
- `attack_type_name`: str — human-readable name
- `nodes`: list of `{id, features, node_y, node_attack_type, node_attack_type_name}`
- `links`: list of `{source, target, edge_features}`
- `num_nodes`, `num_edges`: int
- `id_entropy`, `stats`: additional metadata

### metrics/*.json

Per-configuration evaluation results. 18 files covering 6 datasets x 3 configs (large, small, small_kd).

Each file: `{schema_version, exported_at, data: [{model, scenario, metric_name, value}]}`

Currently only contains `val` scenario — cross-dataset test scenarios are not yet exported.

### Other files

| File | Description |
|---|---|
| `leaderboard.json` | Cross-dataset model comparison (all metrics, all runs) |
| `model_sizes.json` | Parameter counts per model type and scale |
| `training_curves.parquet` | Loss/accuracy curves over training epochs |
| `graph_statistics.parquet` | Per-graph structural statistics |
| `datasets.json` | Dataset metadata |
| `runs.json` | Run metadata |
| `kd_transfer.json` | Knowledge distillation transfer metrics |

## Run ID Convention

Format: `{dataset}/{eval_config}`

- Datasets: `hcrl_sa`, `hcrl_ch`, `set_01` through `set_04`
- Configs: `eval_large_evaluation`, `eval_small_evaluation`, `eval_small_evaluation_kd`

The paper defaults to `hcrl_sa/eval_large_evaluation` for main results and `hcrl_sa/eval_small_evaluation_kd` for CKA.

## Known Gaps

These files need richer exports from the KD-GAT pipeline:

1. **dqn_policy.parquet** — needs per-graph `label` + `attack_type` columns
2. **recon_errors.parquet** — needs per-component error decomposition
3. **attention_weights.parquet** — needs per-edge weights + node positions
4. **metrics/*.json** — needs cross-dataset test scenario results

Until these are addressed, `pull_data.py` preserves existing committed files for the affected figures.