🛡️ SQLGuardian-Transformer

Advanced SQL Injection Detection using Transformer Neural Networks

Deep learning model menggunakan arsitektur Transformer untuk mendeteksi serangan SQL Injection dengan akurasi tinggi. Model ini mampu mengidentifikasi pola-pola serangan kompleks pada query database secara real-time.

📊 Performance Metrics

Metric	Score
Accuracy	99.30%
Precision	99.08%
Recall	99.03%
F1 Score	99.06%
ROC-AUC	99.64%
Balanced Accuracy	99.25%

Confusion Matrix:

True Negative: 3887
False Positive: 21 (0.54%)
False Negative: 22 (0.96%)
True Positive: 2254

🌟 Key Features

✅ High Accuracy: 99.30% detection rate dengan minimal false positives
✅ Transformer Architecture: Multi-head attention mechanism untuk pattern recognition
✅ Real-time Detection: Fast inference (~7-14ms per query)
✅ Comprehensive EDA: 12 visualizations untuk data insights
✅ Production Ready: Optimized threshold dan model serialization
✅ Extensible: Easy to retrain dengan data baru

🗂️ Dataset

Source: Modified_SQL_Dataset.csv

Properties	Details
Total Records	30,921 queries
Normal Queries	19,537 (63.2%)
Attack Queries	11,382 (36.8%)
Features	Query (text), Label (0/1)

Attack Patterns Detected:

OR 1=1 injections
UNION SELECT attacks
Comment injection (--, /*, */)
String concatenation exploits
SQL function injections
Information schema queries

🏗️ Model Architecture

Transformer Configuration

Embedding Dimension: 128
Attention Heads: 4
Feed-Forward Dimension: 256
Transformer Blocks: 2
Dropout Rate: 0.2
Max Sequence Length: 100 tokens
Vocabulary Size: 10,000 words

Layer Structure

Input Layer (100,)
   ↓
Token & Position Embedding (100, 128)
   ↓
Transformer Block 1
   ├─ Multi-Head Attention (4 heads)
   ├─ Layer Normalization
   ├─ Feed Forward Network (256 → 128)
   └─ Residual Connections
   ↓
Transformer Block 2
   ├─ Multi-Head Attention (4 heads)
   ├─ Layer Normalization
   ├─ Feed Forward Network (256 → 128)
   └─ Residual Connections
   ↓
Global Average Pooling
   ↓
Dense Layer (128, ReLU) + Dropout
   ↓
Dense Layer (64, ReLU) + Dropout
   ↓
Output Layer (1, Sigmoid)

Total Parameters: 1,978,113 (7.55 MB)

📦 Installation

Requirements

pip install tensorflow>=2.8.0
pip install scikit-learn>=1.0.0
pip install pandas numpy matplotlib seaborn
pip install joblib tqdm

Clone Repository

git clone https://github.com/yourusername/sqlguardian-transformer.git
cd sqlguardian-transformer

🚀 Quick Start

Training Model

# Load dataset
df = pd.read_csv('Modified_SQL_Dataset.csv')

# Run all cells in notebook sequentially
# Training akan menghasilkan:
# - sql_injection_transformer_model.keras
# - tokenizer.pkl
# - model_config.pkl
# - evaluation_results.pkl

Inference

import tensorflow as tf
import joblib
from tensorflow.keras.preprocessing.sequence import pad_sequences

# Load model dan artifacts
model = tf.keras.models.load_model('sql_injection_transformer_model.keras')
tokenizer = joblib.load('tokenizer.pkl')
config = joblib.load('preprocessing_artifacts.pkl')

# Predict
query = "SELECT * FROM users WHERE id = 1 OR 1=1--"
sequence = tokenizer.texts_to_sequences([query])
padded = pad_sequences(sequence, maxlen=config['max_len'])
probability = model.predict(padded)[0][0]
prediction = "Attack" if probability >= config['optimal_threshold'] else "Normal"

print(f"Prediction: {prediction}")
print(f"Confidence: {probability:.4f}")

📓 Notebook Documentation

Cell-by-Cell Breakdown

CELL 1: Import Libraries

Purpose: Import semua library yang diperlukan untuk data processing, modeling, dan evaluation

Input: None

Output: Libraries loaded

Data processing: pandas, numpy
Visualization: matplotlib, seaborn
ML framework: scikit-learn
Deep learning: tensorflow, keras
Utils: joblib, tqdm, re

CELL 2: Configuration Setup

Purpose: Set random seeds, display options, dan plotting style untuk reproducibility

Input: None

Output:

Random seed: 42
Warnings suppressed
Plot style: seaborn-v0_8-darkgrid
Pandas display optimized

CELL 3: Load Dataset

Purpose: Load dataset SQL injection dari CSV file

Input: Modified_SQL_Dataset.csv

Output: DataFrame dengan 2 kolom

Query: SQL query string
Label: 0 (Normal), 1 (Attack)
Shape: (30921, 2)

Sample Output:

                                    Query  Label
0    " or pg_sleep  (  __TIME__  )  --      1
1  create user name identified by pass...   1

CELL 4: Basic Data Exploration

Purpose: Exploratory analysis dasar untuk memahami struktur dan distribusi data

Input: DataFrame df

Output: Statistical summary

Dataset Shape: (30921, 2)
Column Names: ['Query', 'Label']
Missing Values: 0
Duplicate Rows: 165

Label Distribution:
0    19537  (63.2%)
1    11382  (36.8%)

CELL 5: Feature Engineering

Purpose: Ekstraksi fitur tambahan dari query untuk analisis mendalam

Input: DataFrame dengan kolom Query

Output: 4 fitur baru

Query_Length: Panjang karakter query
Word_Count: Jumlah kata dalam query
Special_Char_Count: Jumlah karakter spesial
Digit_Count: Jumlah digit dalam query

Statistics:

Query_Length: mean=117.4, max=5000
Word_Count: mean=24.8, max=200
Special_Char_Count: mean=16.7, max=600
Digit_Count: mean=8.2, max=80

CELL 6: Comprehensive Visualization

Purpose: Visualisasi 12 subplot untuk memahami karakteristik data dan pola serangan

Input: DataFrame dengan semua fitur

Output: Figure dengan 12 visualizations

Subplots:

Distribusi Label (Countplot): Normal 19,537 vs Attack 11,382
Proporsi Label (Pie Chart): Normal 63.2% vs Attack 36.8%
Distribusi Panjang Query (Violin Plot): Attack cenderung lebih panjang
Distribusi Jumlah Kata (Box Plot): Attack memiliki variance lebih tinggi
Histogram Panjang Query: Separasi jelas antara Normal dan Attack
Histogram Jumlah Kata: Attack lebih spread out
Distribusi Karakter Spesial (Violin Plot): Attack signifikan lebih banyak
Distribusi Angka (Violin Plot): Attack menggunakan lebih banyak digit
Correlation Heatmap: Strong correlation (0.78-0.95) antar fitur
Rata-rata Panjang Query per Label: Attack ~40 chars, Normal ~118 chars
Rata-rata Karakter Spesial per Label: Attack ~17, Normal ~3
Sample Attack Queries: Menampilkan 5 contoh attack patterns

Key Insights:

Attack queries memiliki pola yang distinct
Special characters adalah strong indicator
Query length dapat membantu klasifikasi

CELL 7: Text Preprocessing

Purpose: Clean dan normalize text untuk tokenization

Input: Raw query strings

Output: Cleaned query strings

def clean_text(text):
    - Convert to lowercase
    - Remove multiple whitespaces
    - Strip leading/trailing spaces

Example:

Original: "SELECT * FROM users WHERE id = 1 OR 1=1--"
Cleaned:  "select * from users where id = 1 or 1=1--"

CELL 8: Train-Validation-Test Split

Purpose: Split data dengan stratified sampling untuk mempertahankan proporsi label

Input: X (queries), y (labels)

Output: 3 datasets

Training Set:   21,644 samples (70%)
Validation Set:  3,092 samples (10%)
Test Set:        6,184 samples (20%)

Label Distribution (preserved):
Train:  Normal 13,675 | Attack 7,969
Val:    Normal  1,954 | Attack 1,138
Test:   Normal  3,908 | Attack 2,276

CELL 9: Tokenization and Padding

Purpose: Convert text ke sequences dan pad ke fixed length

Input: Text queries

Configuration:

MAX_WORDS = 10000
MAX_LEN = 100

Output: Padded sequences

Vocabulary Size: 9,847 unique words
Training Shape: (21644, 100)
Validation Shape: (3092, 100)
Test Shape: (6184, 100)

Sample Tokenized:
Original: "select * from users where id = 1"
Tokens: [12, 4, 56, 89, 234, 567, 2, 8]
Padded: [12, 4, 56, 89, 234, 567, 2, 8, 0, 0, ..., 0]

CELL 10: TransformerBlock Layer

Purpose: Define custom Transformer block dengan multi-head attention

Input: None (class definition)

Output: Custom Keras Layer

Components:

Multi-Head Attention (4 heads, key_dim=128)
Feed-Forward Network (256 → 128)
Layer Normalization (2 layers)
Dropout (rate=0.1)
Residual connections

Methods:

call(): Forward pass
get_config(): Serialization
from_config(): Deserialization

CELL 11: TokenAndPositionEmbedding Layer

Purpose: Define embedding layer yang combine token embedding dan positional encoding

Input: None (class definition)

Output: Custom Keras Layer

Components:

Token Embedding (vocab_size → embed_dim)
Position Embedding (maxlen → embed_dim)
Addition operation

Formula: output = token_emb(x) + pos_emb(positions)

CELL 12: Create Model Function

Purpose: Build complete Transformer model architecture

Input: Hyperparameters

Output: Compiled Keras Model

Architecture Summary:

Model: "functional"
_________________________________________________________________
Layer (type)                Output Shape              Param #   
=================================================================
input_layer                 (None, 100)               0         
token_and_position_embedding (None, 100, 128)         1,292,800
transformer_block_1         (None, 100, 128)          330,240
transformer_block_2         (None, 100, 128)          330,240
global_average_pooling      (None, 128)               0         
dropout                     (None, 128)               0         
dense_1                     (None, 128)               16,512
dropout                     (None, 128)               0         
dense_2                     (None, 64)                8,256
dropout                     (None, 64)                0         
dense_output                (None, 1)                 65        
=================================================================
Total params: 1,978,113 (7.55 MB)
Trainable params: 1,978,113 (7.55 MB)
Non-trainable params: 0 (0.00 B)

CELL 13: Build Model

Purpose: Instantiate model dengan specified hyperparameters

Input: Hyperparameters

VOCAB_SIZE = 10000
EMBED_DIM = 128
NUM_HEADS = 4
FF_DIM = 256
NUM_TRANSFORMER_BLOCKS = 2
DROPOUT_RATE = 0.2

Output: Model instance ready for training

CELL 14: Calculate Class Weights

Purpose: Handle imbalanced dataset dengan class weighting

Input: y_train labels

Output: Class weights dictionary

Class Weights: {
    0: 0.7854,  # Normal queries (downweight)
    1: 1.3591   # Attack queries (upweight)
}

Formula: weight = n_samples / (n_classes * n_samples_class)

CELL 15: Define Custom Metrics

Purpose: Create custom F1 score dan balanced accuracy metrics untuk monitoring

Input: None (function definitions)

Output: 2 custom metric functions

f1_metric: Harmonic mean of precision and recall
balanced_accuracy: Average of sensitivity and specificity

Mathematical Formulas:

F1 = 2 * (precision * recall) / (precision + recall)
Balanced Acc = (TPR + TNR) / 2

CELL 16: Compile Model

Purpose: Configure model untuk training dengan optimizer, loss, dan metrics

Input: Model instance

Configuration:

Optimizer: Adam(lr=0.001)
Loss: binary_crossentropy
Metrics: 
  - accuracy
  - precision
  - recall
  - auc
  - f1_metric
  - balanced_accuracy

Output: Compiled model ready untuk fit()

CELL 17: Define Callbacks

Purpose: Setup callbacks untuk mengontrol training process

Input: None

Output: 3 callback objects

Callbacks:

EarlyStopping:
- Monitor: val_auc
- Patience: 15 epochs
- Mode: maximize
- Restore best weights: True
ReduceLROnPlateau:
- Monitor: val_auc
- Factor: 0.5
- Patience: 5 epochs
- Min LR: 1e-7
ModelCheckpoint:
- Save best model to: best_transformer_model.keras
- Monitor: val_auc
- Save best only: True

CELL 18: Train Model

Purpose: Train Transformer model dengan training data

Input:

X_train_pad: (21644, 100)
y_train: (21644,)
X_val_pad: (3092, 100)
y_val: (3092,)

Training Configuration:

Epochs: 50 (early stopped at 17)
Batch Size: 64
Class Weight: Applied
Callbacks: 3 callbacks active
GPU: T4 (enabled)

Output: Training history

Training Progress:

Epoch 1/50: val_auc=0.9964 (BEST) - 34s
Epoch 2/50: val_auc=0.9979 (BEST) - 8s
Epoch 3/50: val_auc=0.9963 - 10s
...
Epoch 7/50: LR reduced to 0.0005
Epoch 12/50: LR reduced to 0.00025
Epoch 17/50: LR reduced to 0.000125, Early Stopping triggered

Best Model: Epoch 2 with val_auc=0.9979

Final Metrics (Epoch 2):

Training: acc=0.9903, auc=0.9949, loss=0.0893
Validation: acc=0.9955, auc=0.9979, loss=0.0434

CELL 19: Plot Training History

Purpose: Visualize training progress dengan 9 subplot metrics

Input: Training history object

Output: Figure dengan 9 subplots

Plots:

Model Loss: Training vs Validation loss convergence
Model Accuracy: 99%+ accuracy dari epoch 2
Model AUC: Near perfect AUC score (0.998)
Model Precision: Consistent 99%+ precision
Model Recall: High recall throughout training
Model F1 Score: Stable F1 around 0.20 (metric issue)
Model Balanced Accuracy: Consistent 0.27 (metric issue)
Learning Rate Schedule: Stepped reduction visible
Final Metrics Comparison: Bar chart train vs val

Key Observations:

Fast convergence (epoch 2)
No overfitting (train ≈ val)
Learning rate reduction effective
Custom metrics have implementation issue (F1, Balanced Acc)

CELL 20: Find Optimal Threshold

Purpose: Determine threshold yang maximize F1 score pada validation set

Input:

y_val: True labels
y_pred_proba_val: Predicted probabilities

Process:

Generate precision-recall curve
Calculate F1 scores untuk semua thresholds
Find threshold dengan F1 maksimum

Output:

Optimal Threshold: 0.3376
F1 Score at Optimal: 0.9943

Why Not 0.5?: Default threshold 0.5 tidak optimal untuk imbalanced dataset. Threshold 0.338 memberikan balance terbaik antara precision dan recall.

CELL 21: Comprehensive Evaluation Function

Purpose: Define function untuk evaluate model dengan multiple metrics

Input: y_true, y_pred, y_pred_proba, dataset_name

Output: Metrics dictionary dan confusion matrix

Metrics Calculated:

Accuracy
Precision
Recall
F1 Score
ROC-AUC
Average Precision
Sensitivity (TPR)
Specificity (TNR)
Balanced Accuracy

Console Output Format:

============================================================
VALIDATION SET EVALUATION RESULTS
============================================================
Accuracy:           0.9958
Precision:          0.9947
Recall:             0.9938
F1 Score:           0.9943
ROC-AUC:            0.9982
Average Precision:  0.9982
Sensitivity:        0.9938
Specificity:        0.9969
Balanced Accuracy:  0.9954
============================================================

Confusion Matrix:
TN: 1948  FP: 6
FN: 7     TP: 1131

CELL 22: Validation and Test Evaluation

Purpose: Run comprehensive evaluation pada validation dan test sets

Input:

Validation: 3,092 samples
Test: 6,184 samples

Output: 2 evaluation results

Validation Set Results:

Accuracy:           99.58%
Precision:          99.47%
Recall:             99.38%
F1 Score:           99.43%
ROC-AUC:            99.82%
Balanced Accuracy:  99.54%

Confusion Matrix:
TN: 1948 (63.0%) | FP: 6 (0.2%)
FN: 7 (0.2%)     | TP: 1131 (36.6%)

Test Set Results:

Accuracy:           99.30%
Precision:          99.08%
Recall:             99.03%
F1 Score:           99.06%
ROC-AUC:            99.64%
Balanced Accuracy:  99.25%

Confusion Matrix:
TN: 3887 (62.9%) | FP: 21 (0.3%)
FN: 22 (0.4%)    | TP: 2254 (36.4%)

Performance Comparison:

Validation slightly better than test (expected)
Minimal degradation (~0.3% accuracy drop)
Excellent generalization

CELL 23: Plot Confusion Matrices

Purpose: Visualize confusion matrices dan metrics comparison

Input: val_cm, test_cm, metrics dictionaries

Output: Figure dengan 3 subplots

Subplots:

Confusion Matrix - Validation Set:
- Heatmap dengan annotations
- Percentage labels
- Color: Blues
Confusion Matrix - Test Set:
- Heatmap dengan annotations
- Percentage labels
- Color: Greens
Validation vs Test Metrics:
- Grouped bar chart
- 6 metrics compared
- Value labels on bars

Insights:

Low false positive rate (0.3-0.5%)
Low false negative rate (0.2-0.4%)
Consistent performance across datasets

CELL 24: ROC and PR Curves

Purpose: Visualize model discrimination ability dengan ROC dan PR curves

Input: True labels dan predicted probabilities

Output: Figure dengan 3 subplots

Subplots:

ROC Curve:
- Validation AUC: 0.9982
- Test AUC: 0.9964
- Both curves near perfect (top-left corner)
- Random classifier baseline shown
Precision-Recall Curve:
- Validation AP: 0.9982
- Test AP: 0.9960
- High precision maintained across recall levels
- Near rectangular shape (ideal)
Prediction Probability Distribution:
- Normal queries: Concentrated at 0.0-0.2
- Attack queries: Concentrated at 0.8-1.0
- Clear separation between classes
- Optimal threshold at 0.338

Interpretation:

Excellent class separation
Model very confident in predictions
Minimal overlap between distributions

CELL 25: Classification Reports

Purpose: Generate detailed sklearn classification reports

Input: y_true, y_pred for validation and test sets

Output: Formatted classification reports

Validation Set Report:

              precision    recall  f1-score   support

      Normal     0.9964    0.9969    0.9967      1954
      Attack     0.9947    0.9938    0.9943      1138

    accuracy                         0.9958      3092
   macro avg     0.9956    0.9954    0.9955      3092
weighted avg     0.9958    0.9958    0.9958      3092

Test Set Report:

              precision    recall  f1-score   support

      Normal     0.9944    0.9946    0.9945      3908
      Attack     0.9908    0.9903    0.9906      2276

    accuracy                         0.9930      6184
   macro avg     0.9926    0.9925    0.9925      6184
weighted avg     0.9930    0.9930    0.9930      6184

CELL 26: Sample Predictions

Purpose: Show random sample predictions untuk qualitative evaluation

Input: 10 random test samples

Output: DataFrame dengan predictions

Sample Output:

Query                                         True_Label  Predicted_Label  Probability  Correct
select * from users where id = 1              Normal      Normal           0.0234       ✓
select * from users where id = 1 or 1=1--     Attack      Attack           0.9876       ✓
admin' or '1'='1                              Attack      Attack           0.9654       ✓
select name from customers                    Normal      Normal           0.0456       ✓
union select null, table_name from info...    Attack      Attack           0.9923       ✓
...

Statistics:

Correct predictions: 10/10 (100%)
High confidence on correct predictions
Clear probability separation

CELL 27: Save Model and Artifacts

Purpose: Save all model components untuk deployment

Input: Model, tokenizer, configs, results

Output: 5 files saved

Files Saved:

sql_injection_transformer_model.keras (7.55 MB)
- Complete model architecture
- Trained weights
- Optimizer state
tokenizer.pkl
- Fitted tokenizer object
- Word index dictionary
- Configuration
model_config.pkl
- Hyperparameters
- Architecture settings
- Optimal threshold
evaluation_results.pkl
- All metrics
- Confusion matrices
- Training history
preprocessing_artifacts.pkl
- MAX_WORDS, MAX_LEN
- Optimal threshold
- Other preprocessing params

CELL 28: Prediction Function

Purpose: Create reusable function untuk inference

Input: SQL queries (string or list)

Output: Prediction results

Function Signature:

def predict_sql_injection(
    queries, 
    model_path='sql_injection_transformer_model.keras',
    tokenizer_path='tokenizer.pkl',
    config_path='preprocessing_artifacts.pkl'
):

Returns: List of dictionaries

[
    {
        'query': 'SELECT * FROM users...',
        'prediction': 'Attack',
        'probability': 0.9876,
        'confidence': 0.9876
    },
    ...
]

Features:

Automatic model loading
Batch prediction support
Confidence score calculation
Truncation for long queries

CELL 29: Test Prediction Function

Purpose: Validate prediction function dengan test queries

Input: 6 test queries (3 normal, 3 attack)

Test Queries:

test_queries = [
    "SELECT * FROM users WHERE id = 1",
    "SELECT * FROM users WHERE id = 1 OR 1=1--",
    "SELECT * FROM products WHERE category = 'electronics'",
    "admin' OR '1'='1",
    "SELECT name, email FROM customers WHERE active = 1",
    "1' UNION SELECT NULL, table_name FROM information_schema.tables--"
]

Output: Predictions untuk semua queries

Query: SELECT * FROM users WHERE id = 1
Prediction: Normal
Probability: 0.0234
Confidence: 0.9766
--------------------------------------------------------------------------------
Query: SELECT * FROM users WHERE id = 1 OR 1=1--
Prediction: Attack
Probability: 0.9876
Confidence: 0.9876
--------------------------------------------------------------------------------
...

Validation: 6/6 correct predictions (100%)

CELL 30: Final Summary

Purpose: Display comprehensive final summary

Input: All evaluation results

Output: Formatted summary

================================================================================
SQL INJECTION DETECTION MODEL TRAINING COMPLETED
================================================================================

Final Test Set Performance:
  Accuracy:          0.9930
  Precision:         0.9908
  Recall:            0.9903
  F1 Score:          0.9906
  ROC-AUC:           0.9964
  Balanced Accuracy: 0.9925

Model Details:
  Total Parameters:  1,978,113
  Model Size:        7.55 MB
  Training Time:     ~8 seconds/epoch
  Best Epoch:        2/50
  Optimal Threshold: 0.3376

Files Generated:
  ✓ sql_injection_transformer_model.keras
  ✓ tokenizer.pkl
  ✓ model_config.pkl
  ✓ evaluation_results.pkl
  ✓ preprocessing_artifacts.pkl

================================================================================

🔬 Technical Details

Hyperparameters

Parameter	Value	Description
Vocabulary Size	10,000	Maximum unique words
Max Sequence Length	100	Padded query length
Embedding Dimension	128	Token embedding size
Attention Heads	4	Multi-head attention
Feed-Forward Dimension	256	FFN hidden layer
Transformer Blocks	2	Number of layers
Dropout Rate	0.2	Regularization
Batch Size	64	Training batch size
Learning Rate	0.001 → 0.000125	Adaptive with ReduceLR
Optimizer	Adam	Gradient descent

Training Configuration

Early Stopping: Patience 15, monitor val_auc
Learning Rate Reduction: Factor 0.5, patience 5
Class Weights: {0: 0.7854, 1: 1.3591}
GPU: NVIDIA T4 (Google Colab)
Training Time: ~8 seconds/epoch
Best Model: Epoch 2

📈 Results Analysis

Strengths

✅ Excellent Detection Rate

99.03% recall (hanya 22 dari 2276 attack yang lolos)
99.08% precision (minimal false alarms)

✅ Fast Convergence

Optimal model achieved di epoch 2
Total training < 5 minutes

✅ No Overfitting

Train dan validation metrics hampir identik
Generalization sangat baik

✅ Production Ready

Fast inference (7-14ms per query)
Optimized threshold (0.338)
Complete serialization

Limitations

⚠️ Dataset Bias

Synthetic dataset dengan pola obvious
Real-world attack mungkin lebih sophisticated

⚠️ Custom Metrics Issue

F1 dan Balanced Accuracy stuck di ~0.20-0.27
Tidak mempengaruhi training (loss-based)

⚠️ Adversarial Robustness

Belum tested dengan obfuscated attacks
Perlu testing dengan encoded payloads (hex, base64)

🙏 Acknowledgments

Dataset: Modified SQL Injection Dataset
Framework: TensorFlow/Keras
Architecture: Transformer (Vaswani et al., 2017)
Platform: Google Colab with T4 GPU

📚 References

Vaswani, A., et al. (2017). "Attention Is All You Need." NeurIPS.
Devlin, J., et al. (2018). "BERT: Pre-training of Deep Bidirectional Transformers." NAACL.
OWASP SQL Injection Prevention Cheat Sheet
SQL Injection Attack Detection Methods - IEEE

Name		Name	Last commit message	Last commit date
Latest commit History 5 Commits
Data		Data
Main		Main
Notebook		Notebook
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
requirements.txt		requirements.txt

License

Arfazrll/SQLGuard-Transformer

Folders and files

Latest commit

History

Repository files navigation

🛡️ SQLGuardian-Transformer

Advanced SQL Injection Detection using Transformer Neural Networks

📊 Performance Metrics

🌟 Key Features

🗂️ Dataset

🏗️ Model Architecture

Transformer Configuration

Layer Structure

📦 Installation

Requirements

Clone Repository

🚀 Quick Start

Training Model

Inference

📓 Notebook Documentation

Cell-by-Cell Breakdown

CELL 1: Import Libraries

CELL 2: Configuration Setup

CELL 3: Load Dataset

CELL 4: Basic Data Exploration

CELL 5: Feature Engineering

CELL 6: Comprehensive Visualization

CELL 7: Text Preprocessing

CELL 8: Train-Validation-Test Split

CELL 9: Tokenization and Padding

CELL 10: TransformerBlock Layer

CELL 11: TokenAndPositionEmbedding Layer

CELL 12: Create Model Function

CELL 13: Build Model

CELL 14: Calculate Class Weights

CELL 15: Define Custom Metrics

CELL 16: Compile Model

CELL 17: Define Callbacks

CELL 18: Train Model

CELL 19: Plot Training History

CELL 20: Find Optimal Threshold

CELL 21: Comprehensive Evaluation Function

CELL 22: Validation and Test Evaluation

CELL 23: Plot Confusion Matrices

CELL 24: ROC and PR Curves

CELL 25: Classification Reports

CELL 26: Sample Predictions

CELL 27: Save Model and Artifacts

CELL 28: Prediction Function

CELL 29: Test Prediction Function

CELL 30: Final Summary

🔬 Technical Details

Hyperparameters

Training Configuration

📈 Results Analysis

Strengths

Limitations

🙏 Acknowledgments

📚 References

About

Topics

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Languages

Packages