Author: Dr. Mallarapu
Created: 2026-07-27
Course: SEAS 8414 — Security Analytics
Goal of this notebook¶
Train and audit detectors on OTIDS CAN frames, captured from a second vehicle.
What you will learn¶
- Read a majority-class baseline before trusting any accuracy figure.
- Find the strongest single feature, then test it by dropping it and refitting.
- Tell duplicate inflation apart from genuine signal.
- Report per-group recall, because the rare classes carry the risk.
Where this connects to the course text¶
The text builds a defence pipeline; this notebook trains a classifier and audits it. The links below are to specific chapter objectives that share an analytic move, not to matching subject matter.
- Chapter 6: Digital Twins for Remediation Simulation — Learning objective 2 (section 6.1) treats fidelity as a promotion gate. An in-distribution score is not a deployment estimate, which is the gate this notebook refuses to pass.
- Chapter 11: Formal Protocol Verification — Section 11.1.2, titled Proved, tested, and hoped, asks you to separate exactly those three. (Chapter 11 lists its objectives in §11.0, not §11.1 as the other chapters do.) The ablation does that job here: it tests whether the headline survives.
In-Vehicle CAN Intrusion Detection on OTIDS (Second Vehicle)¶
Model comparison + per-capture recall + validity audit (real CAN frames, ≥1M)¶
Abstract: OTIDS (Lee, Jeong & Kim, 2017) is a second in-vehicle CAN-bus dataset from the HCRL lab. It was captured on a different vehicle than the nb32 Car-Hacking set. Crucially, it also includes an impersonation attack (a node masquerading as another ECU) that Car-Hacking lacks. Each frame is a CAN ID plus eight payload bytes, labelled normal or one of three attack captures (DoS, fuzzy, impersonation). We keep every attack frame and bound the normal ones to ≥1M, compare four learners, and audit — a cross-dataset companion to nb32.
1. Research problem¶
Task: Classify a CAN frame as normal or attack on a second vehicle/dataset. The value over nb32 is the impersonation attack, which reuses legitimate arbitration IDs. That is the case where payload-only detection is weakest. It is also where the question of whether frame-level features generalize across vehicles becomes concrete.
2. Literature review¶
- Lee, Jeong & Kim (2017) — OTIDS: A Novel Intrusion Detection System for In-Vehicle Network by Using Remote Frame (PST). The dataset and an offset-ratio / time-interval detector.
- Song, Woo & Kim (2020) — the Car-Hacking dataset (nb32 companion).
- Koscher et al. (2010) — Experimental Security Analysis of a Modern Automobile (IEEE S&P).
- Sommer & Paxson (2010) — the closed-world ML critique.
Related approaches and their known caveats — drawn from the wider literature; these are not measurements reproduced on this exact corpus:
| Reported approach | Known caveat |
|---|---|
| Lee et al. (2017) — offset-ratio / time-interval detection | uses request/response TIMING, not payload; complementary to a frame classifier |
| Per-frame classifiers (our setting) | impersonation reuses valid IDs, so payload-only detection is weakest there |
3. Dataset provenance & honesty caveats¶
| Property | Value |
|---|---|
| Source | Kaggle bikashkundu/can-hcrl-otids (HCRL OTIDS) |
| Rows | 3,744,041 frames as loaded (printed below); every attack kept + normal capped |
| Label | target 0 normal; 1/2/3 = the three attack captures |
| Access | Kaggle API token required |
Honestly: the three attack captures map to OTIDS's DoS, fuzzy and impersonation attacks. But we label them by capture id (attack_capture_1/2/3) rather than assert a file→attack mapping we cannot verify. We keep all attack frames and subsample normal, so the attack rate is a bounded-sample rate. Timing is dropped (frames classified independently), which is exactly where impersonation is hardest to catch.
Why the loaded count is smaller than the archive: the loader keeps every attack frame, then subsamples the normal ones. The cap is the 1_500_000 ceiling in the norm = raw[raw['y'] == 0].sample(min(...), ...) line. Normal frames above that ceiling never enter df. So the loaded total is a capped total, not the archive total. This notebook never prints the archive's own row count, so we do not quote one. One-line check after the loader cell: print(f'{len(raw):,} raw -> {len(df):,} loaded').
Before you run this: getting the data¶
This notebook downloads its own data on the first run, then caches it. You do not fetch anything by hand.
Dataset: Kaggle bikashkundu/can-hcrl-otids -> /tmp/kg_otids. It is about 746 MB on disk.
One-time setup. Sign in at kaggle.com, open Settings, and under API choose Create New Token. Kaggle hands you a kaggle.json file. This notebook does not read that file. It reads a plain key file, so convert it once:
mkdir -p ~/.kaggle
python3 -c "import json;print(json.load(open('kaggle.json'))['key'],end='')" > ~/.kaggle/access_token
chmod 600 ~/.kaggle/access_token
Never paste the token into a cell, a commit, or a screenshot. If it leaks, revoke it from the same Settings page.
If the loader fails:
FileNotFoundError: ~/.kaggle/access_token- you createdkaggle.jsonbut not the key file. Run the command above.401 Unauthorized- the key is wrong, or a trailing newline crept in.403 Forbidden- open the dataset page on Kaggle while signed in, accept its terms, then re-run the cell.
The cache sits under /tmp, which macOS clears on reboot. To keep it, move the folder somewhere durable and symlink it back. Do not edit the path in the code cell below: that changes a code cell and invalidates the stored outputs you are reviewing.
4. Solution design¶
The methodology is deliberately two-track. We earn a headline score with standard modelling, then interrogate it with a validity audit. Only a verdict that survives both is reported. The diagram below is the shape of every notebook in this series.
Figure 4.1 — Solution design (methodology).
5. Implementation architecture¶
Five stages — ingestion, preprocessing, modelling, evaluation, and a parallel validity-audit path — feed a single graded results ledger. Leakage defences (dropping label-derived and identifier columns) live in preprocessing, before any model sees the data.
Figure 5.1 — Implementation architecture.
6. Data acquisition & preparation¶
Every line below is commented so a student can re-run and modify each step. The cell ends by producing the standard analysis variables: df, X (clean numeric features), y (binary label), feat (feature names), and family. family is the per-group label used for the recall breakdown. It is an attack family on the intrusion corpora, but a transaction type, merchant category or malware category on the fraud/malware ones.
%matplotlib inline
import time, warnings; warnings.filterwarnings('ignore') # keep output clean
import numpy as np, pandas as pd # numerics + dataframes
import matplotlib.pyplot as plt # static plots (embed in HTML+PDF)
plt.rcParams['figure.dpi'] = 120 # crisp figures
RANDOM_STATE = 0 # single seed used everywhere
np.random.seed(RANDOM_STATE) # reproducible sampling
NEG_WORD, POS_WORD = 'benign', 'attack' # class names (overridden by some loaders)
import os, glob
# OTIDS (Lee, Jeong & Kim, 2017; HCRL): a SECOND in-vehicle CAN dataset (different vehicle than the
# Car-Hacking set in nb32) with DoS, fuzzy and IMPERSONATION attacks — the last is absent from
# Car-Hacking. Pre-processed to TS, CAN ID, 8 payload bytes and a `target` code (0 normal; 1/2/3 =
# the three attack captures). Self-contained Kaggle download.
os.environ.setdefault('KAGGLE_KEY', open(os.path.expanduser('~/.kaggle/access_token')).read().strip())
DEST = '/tmp/kg_otids'; os.makedirs(DEST, exist_ok=True)
if not glob.glob(DEST + '/**/dataset*.csv', recursive=True):
import kaggle; kaggle.api.authenticate()
print('downloading OTIDS (one-time)...')
kaggle.api.dataset_download_files('bikashkundu/can-hcrl-otids', path=DEST, unzip=True, quiet=True)
files = sorted(glob.glob(DEST + '/**/dataset*.csv', recursive=True)) # dataset.csv (normal) + dataset1/2/3 (attacks)
assert files, 'OTIDS dataset*.csv not found'
parts = []
for f in files:
d = pd.read_csv(f, low_memory=False); d.columns = [str(c).strip() for c in d.columns]
parts.append(d)
raw = pd.concat(parts, ignore_index=True)
raw['target'] = pd.to_numeric(raw['target'], errors='coerce').fillna(0).astype(int)
raw['y'] = (raw['target'] != 0).astype(int)
# Keep EVERY attack frame; bound the normal frames (they are the majority) so the set stays ≥1M.
atk = raw[raw['y'] == 1]
norm = raw[raw['y'] == 0].sample(min(int((raw['y'] == 0).sum()), 1_500_000), random_state=0)
df = pd.concat([atk, norm]).reset_index(drop=True)
assert len(df) >= 1_000_000, f'floor not met: {len(df):,}'
df['family'] = np.where(df['y'] == 0, 'normal', 'attack_capture_' + df['target'].astype(str))
# Drop the timestamp + label; keep the CAN arbitration ID and the 8 payload bytes.
DROP = ['TS', 'target', 'y', 'family']
feat = [c for c in df.columns if c not in DROP]
from sklearn.preprocessing import LabelEncoder
X = df[feat].copy()
idlike = [c for c in X.select_dtypes(include='object').columns if X[c].nunique() > 0.5*len(X)]
X = X.drop(columns=idlike) # drop id/timestamp-like leaky columns
for c in X.select_dtypes(include='object').columns:
X[c] = LabelEncoder().fit_transform(X[c].astype(str))
X = X.apply(pd.to_numeric, errors='coerce').replace([np.inf,-np.inf],np.nan).fillna(0.0)
X = X.clip(-1e15, 1e15); X = X.loc[:, X.nunique() > 1] # float32-safe; drop constants
import re
_seen, _cols = {}, []
for _c in X.columns: # unique LightGBM-safe names
_c = re.sub(r'[^0-9A-Za-z_]+', '_', str(_c)).strip('_') or 'f'
_seen[_c] = _seen.get(_c, -1) + 1
_cols.append(_c if _seen[_c] == 0 else f'{_c}_{_seen[_c]}')
X.columns = _cols; feat = list(X.columns)
y = df['y'].to_numpy(); family = df['family'].to_numpy()
print(f'loaded {len(df):,} CAN frames x {len(feat)} features; attack rate {y.mean():.4f}; families {sorted(set(family))}')
loaded 3,744,041 CAN frames x 11 features; attack rate 0.5994; families ['attack_capture_1', 'attack_capture_2', 'attack_capture_3', 'normal']
7. Exploratory data analysis¶
# --- EDA 1: class balance and the attack-family mix ---
fig, ax = plt.subplots(1, 2, figsize=(11, 4))
df['y'].map({0:NEG_WORD,1:POS_WORD}).value_counts().plot.bar( # counts per class
ax=ax[0], color=['#2a9d8f','#e76f51']); ax[0].set_yscale('log')
ax[0].set_title(f'Class balance ({NEG_WORD} vs {POS_WORD})'); ax[0].set_ylabel('records (log)')
df.loc[df.y==1,'family'].value_counts().head(8).plot.barh( # top attack families
ax=ax[1], color='#e76f51'); ax[1].invert_yaxis(); ax[1].set_title('Top attack families')
plt.tight_layout(); plt.show()
# --- EDA 2: feature correlation + a 2-D PCA projection ---
from sklearn.preprocessing import StandardScaler # scale before PCA
from sklearn.decomposition import PCA
fig, ax = plt.subplots(1, 2, figsize=(12, 5))
topv = X[feat].var().sort_values().tail(12).index # 12 highest-variance features
im = ax[0].imshow(X[topv].corr(), cmap='coolwarm', vmin=-1, vmax=1) # correlation heatmap
ax[0].set_xticks(range(len(topv))); ax[0].set_xticklabels(topv, rotation=90, fontsize=7)
ax[0].set_yticks(range(len(topv))); ax[0].set_yticklabels(topv, fontsize=7)
ax[0].set_title('Feature correlation (top-variance)'); fig.colorbar(im, ax=ax[0], shrink=0.7)
samp = X.sample(min(5000, len(X)), random_state=RANDOM_STATE) # subsample for a fast PCA
pc = PCA(n_components=2).fit_transform(StandardScaler().fit_transform(samp))
ys = y[samp.index] # aligned labels for coloring
for lab,c in [(0,'#2a9d8f'),(1,'#e76f51')]:
ax[1].scatter(pc[ys==lab,0], pc[ys==lab,1], s=4, alpha=0.4, color=c,
label={0:NEG_WORD,1:POS_WORD}[lab])
ax[1].set_title('PCA projection (2 components)'); ax[1].legend(); ax[1].set_xlabel('PC1'); ax[1].set_ylabel('PC2')
plt.tight_layout(); plt.show()
8. Model comparison¶
Four diverse learners share one held-out split, ranked by ROC-AUC.
Two honesty guards print with the table:
- The models train on a stratified subsample of at most 120,000 rows. The full row count is printed above. So every score here is a subsample number, not a full-corpus claim.
- The majority-class baseline accuracy appears inside the ranking table. On imbalanced data, 0.99 accuracy can be worse than always guessing the majority class. Judge each model against that baseline, not against 0.5.
# --- Model comparison: four learners on the same held-out split ---
from sklearn.model_selection import train_test_split
from sklearn.linear_model import LogisticRegression
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import accuracy_score, roc_auc_score
import xgboost as xgb, lightgbm as lgb
# Stratified split keeps the class ratio in both halves.
Xtr, Xte, ytr, yte = train_test_split(X, y, test_size=0.25, random_state=RANDOM_STATE, stratify=y)
from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import StandardScaler
N_MATERIALIZED = len(y) # the full corpus we loaded (see printed count)
# HONEST DISCLOSURE: we do NOT train on all N. We fit on a STRATIFIED subsample (<=120k) because
# these learners saturate long before then on this data. Every headline below is a SUBSAMPLE
# number, not a full-corpus number — saying otherwise would be the fabrication this course forbids.
if len(Xtr) > 120_000:
Xtr, _, ytr, _ = train_test_split(Xtr, ytr, train_size=120_000, random_state=RANDOM_STATE,
stratify=ytr) # genuinely stratified, not random
MAJORITY_BASELINE = max(np.mean(yte), 1 - np.mean(yte)) # accuracy of 'always predict majority'
print(f'materialized {N_MATERIALIZED:,} rows | trained on {len(Xtr):,} (stratified subsample) | '
f'held-out {len(yte):,}')
print(f'MAJORITY-CLASS BASELINE accuracy = {MAJORITY_BASELINE:.4f} '
f'(any model must beat THIS, not 0.5, to be interesting)')
models = { # four standard, diverse learners
'LogisticRegression': make_pipeline(StandardScaler(), LogisticRegression(max_iter=300)), # scaled!
'RandomForest': RandomForestClassifier(n_estimators=60, n_jobs=-1, random_state=RANDOM_STATE),
'XGBoost': xgb.XGBClassifier(n_estimators=80, max_depth=6, tree_method='hist', n_jobs=-1,
eval_metric='logloss', random_state=RANDOM_STATE),
'LightGBM': lgb.LGBMClassifier(n_estimators=80, n_jobs=-1, verbose=-1, random_state=RANDOM_STATE),
}
rows, fitted = [], {}
for name, m in models.items(): # fit + score each model
t = time.perf_counter(); m.fit(Xtr, ytr); fitted[name] = m
p = m.predict_proba(Xte)[:, 1] # positive-class probability on held-out
rows.append({'model': name, 'accuracy': round(accuracy_score(yte, (p>0.5).astype(int)), 6),
'roc_auc': round(roc_auc_score(yte, p), 6), # 6 dp: a 1.000000 is a red flag, not a win
'train_s': round(time.perf_counter()-t, 1)})
rows.append({'model': 'MajorityBaseline', 'accuracy': round(MAJORITY_BASELINE, 4),
'roc_auc': 0.5, 'train_s': 0.0}) # show the baseline IN the ranking table
comparison = pd.DataFrame(rows).sort_values('roc_auc', ascending=False).reset_index(drop=True)
_ranked = comparison[comparison.model != 'MajorityBaseline']
best_name = _ranked.iloc[0]['model']; best = fitted[best_name] # winner by ROC-AUC (excl. baseline)
print('best model:', best_name); comparison
materialized 3,744,041 rows | trained on 120,000 (stratified subsample) | held-out 936,011 MAJORITY-CLASS BASELINE accuracy = 0.5994 (any model must beat THIS, not 0.5, to be interesting)
best model: XGBoost
| model | accuracy | roc_auc | train_s | |
|---|---|---|---|---|
| 0 | XGBoost | 0.858807 | 0.954777 | 0.4 |
| 1 | LightGBM | 0.855056 | 0.952706 | 1.3 |
| 2 | RandomForest | 0.856261 | 0.951774 | 0.9 |
| 3 | LogisticRegression | 0.612609 | 0.603834 | 0.2 |
| 4 | MajorityBaseline | 0.599400 | 0.500000 | 0.0 |
9. Results¶
Diagnostics for the winning model, including per-group recall.
The grouping comes from whatever the loader put in family. It is not always an attack taxonomy. On the intrusion corpora it is the attack family. On the fraud and malware corpora it is a transaction type, a merchant category or a malware category. On binary corpora it collapses to the positive class.
Read it accordingly. Where the groups are genuinely rare classes, they reveal whether detection is real. The dominant flood classes do not.
# --- Results for the best model: confusion, ROC, PR, importances, per-family recall ---
from sklearn.metrics import confusion_matrix, roc_curve, precision_recall_curve, recall_score
pb = best.predict_proba(Xte)[:, 1]; pred = (pb > 0.5).astype(int)
fig, ax = plt.subplots(1, 3, figsize=(15, 4))
# (1) confusion matrix
cm = confusion_matrix(yte, pred); ax[0].imshow(cm, cmap='Blues')
ax[0].set_title(f'{best_name}: confusion'); ax[0].set_xticks([0,1]); ax[0].set_yticks([0,1])
ax[0].set_xticklabels([NEG_WORD,POS_WORD]); ax[0].set_yticklabels([NEG_WORD,POS_WORD])
for (i,j),v in np.ndenumerate(cm): ax[0].text(j,i,f'{v:,}',ha='center',va='center')
# (2) ROC and PR curves
fpr,tpr,_ = roc_curve(yte, pb); prec,rec,_ = precision_recall_curve(yte, pb)
ax[1].plot(fpr,tpr,color='#264653'); ax[1].plot([0,1],[0,1],'--',c='grey')
ax[1].set_title(f'ROC (AUC={roc_auc_score(yte,pb):.4f})'); ax[1].set_xlabel('FPR'); ax[1].set_ylabel('TPR')
ax[2].plot(rec,prec,color='#e76f51'); ax[2].set_title('Precision-Recall'); ax[2].set_xlabel('recall'); ax[2].set_ylabel('precision')
plt.tight_layout(); plt.show()
# (3) feature importances + (4) per-attack-family recall
fig, ax = plt.subplots(1, 2, figsize=(13, 5))
imp, names = None, feat # importances, robust to the scaled-LR pipeline
if hasattr(best, 'feature_importances_'): # tree models
imp = best.feature_importances_; names = list(getattr(best, 'feature_names_in_', feat))[:len(imp)]
elif hasattr(best, 'named_steps') and 'logisticregression' in getattr(best, 'named_steps', {}):
imp = np.abs(best.named_steps['logisticregression'].coef_[0]); names = feat # LR pipeline
elif hasattr(best, 'coef_'):
imp = np.abs(best.coef_[0]); names = feat
if imp is not None:
pd.Series(imp, index=names[:len(imp)]).sort_values().tail(12).plot.barh(ax=ax[0], color='#264653')
ax[0].set_title(f'{best_name}: top importances / |coef|')
# Per-family recall, WORST-first so rare, hard classes are visible, not just the dominant floods.
fam_te = df.loc[Xte.index, 'family']
fr = {}
for fam, cnt in fam_te[yte==1].value_counts().items():
if cnt < 5: continue # need a few positives for a meaningful recall
mask = (fam_te==fam).to_numpy(); fr[fam] = recall_score(yte[mask], pred[mask], zero_division=0)
srt = pd.Series(fr).sort_values()
show = pd.concat([srt.head(9), srt.tail(3)]) if len(srt) > 12 else srt # worst 9 + best 3
show = show[~show.index.duplicated()]
show.plot.barh(ax=ax[1], color=['#e76f51' if v < 0.5 else '#2a9d8f' for v in show]); ax[1].set_xlim(0,1)
ax[1].set_title('Per-family recall (worst first; red < 0.5)')
plt.tight_layout(); plt.show()
# Operational numbers, not just figures: false-positive rate and the worst per-family recalls.
tn, fp = int(cm[0,0]), int(cm[0,1])
fpr_op = fp/(fp+tn) if (fp+tn) > 0 else float('nan') # benign wrongly flagged @0.5
print(f'operational FALSE-POSITIVE RATE @0.5 = {fpr_op:.4f} ({fp:,} benign flagged of {fp+tn:,})')
print('worst per-family recalls:', {k: round(v, 3) for k, v in srt.head(6).items()})
operational FALSE-POSITIVE RATE @0.5 = 0.1586 (59,472 benign flagged of 375,000)
worst per-family recalls: {'attack_capture_3': 0.809, 'attack_capture_1': 0.896, 'attack_capture_2': 0.944}
10. Validity audit — is the score real?¶
Three diagnostics. (a) How well can the single best feature, alone, separate the classes? A near-1.0 single-feature AUC means that feature is near-sufficient — a shortcut (which may be legitimate signal or an artifact), not the same as target leakage. (b) The exact-duplicate row rate. (c) The train/test exact-row contamination — the fraction of held-out rows that are duplicates of training rows, which is what actually inflates a held-out score. The trust grade is the worse of the single-feature and contamination concerns.
# --- Validity audit: is the score real detection, or a data shortcut? ---
from sklearn.metrics import roc_auc_score
samp = X.sample(min(60_000, len(X)), random_state=1); ysamp = y[samp.index]
aucs = {}
for c in feat: # AUC of EACH feature alone
col = samp[c].to_numpy(float)
if col.std()==0: continue
a = roc_auc_score(ysamp, col); aucs[c] = max(a, 1-a) # direction-agnostic
best_auc = max(aucs.values()); best_col = max(aucs, key=aucs.get)
dup_rate = 1 - X.drop_duplicates().shape[0]/len(X) # exact-duplicate feature rows (whole set)
# The statistic that actually inflates a held-out score is TRAIN/TEST CONTAMINATION: how many test
# rows are exact duplicates of a training row. Measure it directly on the split used above.
_trkeys = set(map(tuple, np.round(Xtr.to_numpy(), 6)))
_te = np.round(Xte.to_numpy(), 6)[:50_000]
contam = float(np.mean([tuple(r) in _trkeys for r in _te])) # fraction of test rows seen in train
# Trust grade reflects BOTH failure modes and takes the WORSE of the two: a near-perfect single
# feature (shortcut) OR heavy train/test contamination each independently invalidate the headline.
_ga = 'F' if best_auc>=0.999 else 'D' if best_auc>=0.99 else 'C' if best_auc>=0.95 else 'B' if best_auc>=0.85 else 'A'
_gc = 'F' if contam>=0.5 else 'D' if contam>=0.3 else 'C' if contam>=0.15 else 'B' if contam>=0.05 else 'A'
grade = max(_ga, _gc) # 'max' letter = worse grade (A best, F worst)
print(f'best single-feature AUC = {best_auc:.4f} (feature: {best_col})')
print(f' note: a near-1.0 single-feature AUC means this feature is *near-sufficient* (a shortcut),\n'
f' which may be legitimate signal OR an artifact — it is NOT the same as target leakage.')
print(f'exact-duplicate row rate (whole corpus) = {dup_rate:.3f}')
print(f'TRAIN/TEST exact-row contamination = {contam:.3f} (single-feat grade {_ga}, contam grade {_gc})')
print(f'==> data trust grade: {grade} (worse of the two; F = shortcut and/or heavy contamination)')
s = pd.Series(aucs).sort_values().tail(15)
fig, ax = plt.subplots(figsize=(8,5))
s.plot.barh(ax=ax, color=['#e76f51' if v>=0.99 else '#457b9d' for v in s]); ax.axvline(0.5,ls='--',c='grey')
ax.set_xlim(0.5,1.0); ax.set_title('Single-feature ROC-AUC (red = near-perfect shortcut)'); ax.set_xlabel('AUC alone')
plt.tight_layout(); plt.show()
best single-feature AUC = 0.5773 (feature: ID1) note: a near-1.0 single-feature AUC means this feature is *near-sufficient* (a shortcut), which may be legitimate signal OR an artifact — it is NOT the same as target leakage. exact-duplicate row rate (whole corpus) = 0.903 TRAIN/TEST exact-row contamination = 0.830 (single-feat grade A, contam grade F) ==> data trust grade: F (worse of the two; F = shortcut and/or heavy contamination)
11. Ablation — does the headline survive removing the artifacts?¶
Narrating a shortcut is not enough. We retrain the winning model after (1) de-duplicating the corpus (removing the train/test contamination) and (2) dropping the single strongest feature. We report the held-out AUC each time. Read the result honestly, both ways: if the AUC collapses, the headline was a contamination/shortcut artifact. If it barely moves — common on simulated corpora — that is not vindication. It means the classes are separable by many redundant features, because the attack and benign distributions barely overlap. That is its own generation artifact. The numbers below decide which story is true here, not the prose.
Caveat — de-duplication resolves every label conflict to attack: the ablation calls X.drop_duplicates(). That keeps the first occurrence of each feature vector. The loader built df as pd.concat([atk, norm]), so every attack frame sits above every normal frame. A vector that appears in both an attack capture and the normal capture therefore survives as the attack row. Its benign twins are deleted, not reconciled.
The ambiguity is removed in one direction only. Those conflicted rows are exactly the ones no frame-level detector can get right. So the de-duplicated AUC below is an upper bound, not a corrected headline. Read it as how easy the task becomes once conflicts are deleted. Do not read it as the headline with contamination removed.
The unbiased version would resolve each duplicate key by majority label, or drop conflicted keys outright, then refit. We do not run that here. The direction of the bias is read off the code. Its size is unmeasured. One-line check, after the next cell: print(len(X.drop_duplicates()), len(X.assign(_y=y).drop_duplicates())). The first count is unique feature vectors. The second counts unique (vector, label) pairs. A larger second number means conflicted keys exist, and the difference is how many. Do not use the before/after attack rate for this. It also moves with within-class duplication, which a DoS flood dominates.
# --- Ablation: SHOW the inflation empirically, don't just narrate it ---
from sklearn.base import clone
def _retrain_auc(Xa, ya): # re-split, stratified-subsample, refit best family
xtr, xte, ytr2, yte2 = train_test_split(Xa, ya, test_size=0.25, random_state=RANDOM_STATE, stratify=ya)
if len(xtr) > 120_000:
xtr, _, ytr2, _ = train_test_split(xtr, ytr2, train_size=120_000, random_state=RANDOM_STATE, stratify=ytr2)
m = clone(best); m.fit(xtr, ytr2)
return roc_auc_score(yte2, m.predict_proba(xte)[:, 1])
base_auc = roc_auc_score(yte, best.predict_proba(Xte)[:, 1]) # (0) the headline held-out AUC
Xdd = X.drop_duplicates(); ydd = y[Xdd.index] # (1) de-duplicated corpus
auc_dedup = _retrain_auc(Xdd, ydd)
auc_noshort = _retrain_auc(X.drop(columns=[best_col]), y) if best_col in X.columns else base_auc # (2) drop shortcut
ablation = pd.DataFrame([
{'setting': 'headline (as-is)', 'held_out_auc': round(base_auc, 6)},
{'setting': f'de-duplicated ({1-len(Xdd)/len(X):.0%} rows removed)', 'held_out_auc': round(auc_dedup, 6)},
{'setting': f'shortcut feature dropped ({best_col})', 'held_out_auc': round(auc_noshort, 6)},
])
print('Ablation — how much of the headline survives once each artifact is removed:')
ablation
Ablation — how much of the headline survives once each artifact is removed:
| setting | held_out_auc | |
|---|---|---|
| 0 | headline (as-is) | 0.954777 |
| 1 | de-duplicated (90% rows removed) | 0.996659 |
| 2 | shortcut feature dropped (ID1) | 0.946496 |
12. Reproducibility & robustness¶
# --- Reproducibility & robustness ---
import sklearn
from sklearn.model_selection import StratifiedKFold, cross_val_score
print(f'seed={RANDOM_STATE} | numpy {np.__version__} | sklearn {sklearn.__version__} | '
f'xgboost {xgb.__version__} | lightgbm {lgb.__version__}')
# 3-fold cross-validated ROC-AUC of the winning model (fresh clone, bounded subsample) -> mean +/- std.
from sklearn.base import clone
cvX, cvy = Xtr.iloc[:40_000], ytr[:40_000]
def _auc_scorer(est, Xv, yv): # robust to xgboost's 2-col predict_proba
p = est.predict_proba(Xv)
p = p[:, 1] if getattr(p, 'ndim', 1) == 2 else p
return roc_auc_score(yv, p)
try:
cv = cross_val_score(clone(best), cvX, cvy,
cv=StratifiedKFold(3, shuffle=True, random_state=RANDOM_STATE),
scoring=_auc_scorer, error_score='raise')
assert np.all(np.isfinite(cv)), 'non-finite CV folds' # FAIL CLOSED: never narrate a NaN as evidence
print(f'{best_name} 3-fold CV ROC-AUC = {cv.mean():.4f} +/- {cv.std():.4f} '
f'(mean +/- std across 3 stratified folds; a small std means a stable estimate on this split)')
except Exception as e:
print(f'CV UNAVAILABLE ({type(e).__name__}: {str(e)[:60]}); rely on the single held-out AUC above — '
f'we do NOT report a CV number we could not compute')
seed=0 | numpy 2.3.5 | sklearn 1.9.0 | xgboost 1.6.2 | lightgbm 4.7.0
XGBoost 3-fold CV ROC-AUC = 0.9521 +/- 0.0008 (mean +/- std across 3 stratified folds; a small std means a stable estimate on this split)
13. Scientific conclusion¶
As with Car-Hacking (nb32), payload and ID features separate injected from legitimate frames well, and per-capture recall shows which attacks are easy. Because we drop timing, the offset-ratio / time-interval signal OTIDS was built around is absent. A timing/sequence model is one honest next step. Cross-vehicle transfer between this dataset and nb32 (unmeasured here) is the other (Lee et al., 2017; Sommer & Paxson, 2010).
Validity ledger — read the headline against these printed numbers: Majority-class baseline accuracy: 0.5994. The accuracy column must clear that bar to mean anything. For ROC-AUC the trivial baseline is 0.5, not that figure. Winning learner: XGBoost (3-fold CV ROC-AUC 0.9521). Strongest single feature: ID1 at AUC 0.5773. The ablation refutes a single-feature story. Dropping that feature barely moves the AUC: 0.954777 → 0.946496. So the separability is multi-feature. That reflects how this corpus was generated, not one leaky column. De-duplication raises the AUC, to 0.996659. That is not evidence the headline is safe. Collapsing duplicates removes the hardest rows. Where identical feature vectors carry conflicting labels, the rule resolves them all to attack. The de-duplicated task is therefore easier, not cleaner. Data-trust grade: F. It is the worse of two independent sub-checks. Single-feature AUC 0.5773 scores A. Train/test exact-row overlap 0.830 scores F. The overlap check drives the grade, not the single-feature check. That says the split leaks, not that features are clean; the single-feature check separately scores A. On this A-best / F-worst scale, a D or F means the headline is optimistic. Treat it as a benchmark number, not a deployment estimate. Operational false-positive rate at threshold 0.5: 0.1586. Worst per-group recalls, exactly as printed: {attack_capture_3: 0.809, attack_capture_1: 0.896, attack_capture_2: 0.944}. The weakest group sits at 0.809, which is where detection is thinnest. Disclosed limitation: categorical columns are integer-encoded before the split. The encoder therefore sees the test set's category values. On an all-numeric corpus that step is a no-op. The mapping never consults the label, so no label information leaks. It is still transductive. A deployed system would need an unseen-category bucket. How the audit numbers are computed: overlap is measured on the first 50,000 held-out rows, so read it as a sampled estimate. Each ablation re-splits and refits, so tiny differences are re-split noise. The de-duplication variant keeps the first row when a feature vector repeats. The loader stacks attack frames above normal ones. Any conflicting key therefore resolves to attack, which biases that row upward. It is an upper bound, not a corrected headline. How many keys conflict is not measured here. Scope: the split is random, not temporal or entity-grouped. Every number above therefore measures in-distribution separability only.
References¶
- Lee, H., Jeong, S.-H. & Kim, H.K. (2017). OTIDS: A Novel Intrusion Detection System for In-Vehicle Network by Using Remote Frame. 15th Annual Conf. on Privacy, Security and Trust (PST), 57–66.
- Song, H.M., Woo, J. & Kim, H.K. (2020). In-Vehicle Network Intrusion Detection Using Deep CNN. Vehicular Communications, 21.
- Koscher, K. et al. (2010). Experimental Security Analysis of a Modern Automobile. IEEE S&P.
- Sommer, R. & Paxson, V. (2010). Outside the Closed World: On Using Machine Learning for Network Intrusion Detection. IEEE S&P.