Add model card

README.md

@@ -0,0 +1,70 @@
+---
+license: unknown
+tags:
+- security
+- sql-injection
+- scikit-learn
+- anomaly-detection
+- text-classification
+pipeline_tag: text-classification
+---
+
+# VNU SQLi Detection Models
+
+Trained model artifacts for a 2-branch AI-based SQL Injection detection system (Branch 3 —
+session-level — is designed but not yet implemented; see the project repo for details).
+
+Data used to train these models: [Jason-42195/VNU-SQLi-Detection](https://huggingface.co/datasets/Jason-42195/VNU-SQLi-Detection).
+
+## `nhanh1_v1/` — Branch 1 (supervised multiclass)
+
+TF-IDF (char_wb, 2-4gram) + Logistic Regression. Classifies a query into one of 5 classes:
+`normal`, `union_based`, `error_based`, `boolean_blind`, `time_blind`.
+
+- **F1-macro: 0.9822** on held-out test set (13,560 rows)
+- p50 latency: ~0.5ms, size: ~3.9MB
+- Files: `vectorizer.joblib` (TfidfVectorizer), `model.joblib` (LogisticRegression), `metadata.json`
+
+```python
+import joblib
+vectorizer = joblib.load("nhanh1_v1/vectorizer.joblib")
+clf = joblib.load("nhanh1_v1/model.joblib")
+
+X = vectorizer.transform(["1' OR '1'='1"])
+clf.predict(X)  # -> array([3])  (3 = boolean_blind)
+```
+
+## `nhanh2_v1/` — Branch 2 (anomaly detection)
+
+One-Class SVM trained on 100% benign traffic, using 4 structural features (length,
+special_char_ratio, sql_keyword_count, entropy) — not TF-IDF, so it can generalize to
+unseen attack syntax.
+
+- **AUC: 0.90**, FPR: 0.3% (9/3000 benign), detection rate: 20.7% (5196/25065 anomalous)
+- Files: `model.joblib` (wraps `src.models.nhanh2_anomaly.AnomalyDetector`), `metadata.json`
+
+```python
+# From within the project repo (needs src.models.nhanh2_anomaly.AnomalyDetector):
+from src.models.nhanh2_anomaly import AnomalyDetector
+import numpy as np
+
+detector = AnomalyDetector.load("nhanh2_v1")
+X = np.array([[40, 0.05, 1, 3.6]])  # [length, special_char_ratio, sql_keyword_count, entropy]
+detector.score(X)          # continuous anomaly score
+detector.anomaly_flags(X)  # boolean flag
+```
+
+## Limitations
+
+- Branch 1's `boolean_blind` class has ~13% measured label noise (catch-all bucket for
+  unmatched attack rows) — see `data_contract.md` in the project repo.
+- Branch 2's detection rate (20.7%) reflects a diverse anomalous eval set (D3 CSIC2010,
+  covering XSS/path-traversal/etc., not just SQLi) — not a pure SQLi zero-day benchmark.
+- No adversarial/obfuscation robustness testing yet.
+- License: mixed/unclear for underlying training data (see the dataset repo's card) —
+  treat as research/course-project artifacts, not yet cleared for unrestricted reuse.
+
+## Project repo
+
+Full source, training scripts, and documentation: see the `VNU-Database2-Project` repo
+(private/course project — ask the author for access).