app.py

import streamlit as st
import tensorflow as tf
from tensorflow.keras.models import load_model
from tensorflow.keras.preprocessing.text import Tokenizer
from tensorflow.keras.preprocessing.sequence import pad_sequences
import pickle
import re
import time
import numpy as np
from sklearn.ensemble import RandomForestClassifier
from sklearn.svm import SVC

# Load models and preprocessing components
@st.cache_resource
def load_components():
    # Load deep learning models
    cnn_model = load_model('cnn_model.h5')
    lstm_model = load_model('lstm_model.h5')
    
    # Load traditional ML models
    with open('rf_model.pkl', 'rb') as f:
        rf_model = pickle.load(f)
    with open('svm_model.pkl', 'rb') as f:
        svm_model = pickle.load(f)
    
    # Load tokenizer and vectorizer
    with open('sql_tokenizer.pkl', 'rb') as f:
        tokenizer_data = pickle.load(f)
    with open('tfidf_vectorizer.pkl', 'rb') as f:
        tfidf_vectorizer = pickle.load(f)
    
    return {
        'cnn_model': cnn_model,
        'lstm_model': lstm_model,
        'rf_model': rf_model,
        'svm_model': svm_model,
        'tokenizer': tokenizer_data['tokenizer'],
        'max_sequence_length': tokenizer_data['max_sequence_length'],
        'tfidf_vectorizer': tfidf_vectorizer
    }

# Try to load all components
try:
    components = load_components()
    model_loading_error = None
except Exception as e:
    model_loading_error = str(e)
    components = None

# Preprocess functions
def preprocess_query_for_deep_learning(query, tokenizer, max_sequence_length):
    """
    Tokenizes and pads the input query to prepare it for deep learning models.
    """
    sequences = tokenizer.texts_to_sequences([query])
    padded = pad_sequences(sequences, maxlen=max_sequence_length, padding='post')
    return padded

def preprocess_query_for_traditional_ml(query, tfidf_vectorizer):
    """
    Transforms the input query using TF-IDF for traditional ML models.
    """
    return tfidf_vectorizer.transform([query])

# Define improved regex patterns for SQL injection attempts
SQL_INJECTION_PATTERNS = [
    # SQL comment syntax that follows a quote (likely injection)
    r"(?i)'.*--",
    
    # Quote followed by OR/AND with comparison (classic injection pattern)
    r"(?i)'\s*(OR|AND)\s*['\d\w]+=\s*['\d\w]+",
    
    # SQL Comment without preceding from a query context
    r"(?i)(\s|^)--",
    
    # Multiple query execution with semicolon
    r"(?i)'.*;.*--",
    
    # UNION-based injections
    r"(?i)'\s*UNION\s+(ALL\s+)?SELECT",
    
    # Time-delay attacks
    r"(?i)'\s*;\s*WAITFOR\s+DELAY",
    
    # DROP/ALTER table attacks
    r"(?i)'\s*;\s*(DROP|ALTER)",
    
    # Quote followed by a true condition
    r"(?i)'\s*OR\s*'?\d+'?\s*=\s*'?\d+'?",
    
    # Quote followed by always true condition like 1=1
    r"(?i)'\s*OR\s*(['\"]\d+['\"])=(['\"]\d+['\"])",
    
    # Batch queries
    r"(?i);\s*(SELECT|INSERT|UPDATE|DELETE|DROP)",
    
    # CAST attacks
    r"(?i)CAST\s*\(.+AS\s+.+\)",
    
    # Typical SQL function calls in injections
    r"(?i)'\s*;\s*(EXEC|EXECUTE).*",
]

# Safe SQL patterns that should not trigger false positives
SAFE_SQL_PATTERNS = [
    # Standard SELECT query
    r"(?i)^SELECT\s+[\w\d\s,*]+\s+FROM\s+[\w\d]+(\s+WHERE\s+[\w\d\s=<>']+)?$",
    
    # Standard INSERT query
    r"(?i)^INSERT\s+INTO\s+[\w\d]+\s*\([^)]+\)\s*VALUES\s*\([^)]+\)$",
    
    # Standard UPDATE query
    r"(?i)^UPDATE\s+[\w\d]+\s+SET\s+[\w\d\s=',]+(\s+WHERE\s+[\w\d\s=<>']+)?$",
]

# Rule-based detection function
def detect_sql_injection_with_regex(query):
    """
    Detects potential SQL injection patterns using improved regex.
    Returns True if any malicious pattern matches and no safe pattern matches.
    """
    # First check if the query matches any safe pattern
    for pattern in SAFE_SQL_PATTERNS:
        if re.search(pattern, query.strip()):
            # Query matches a safe pattern
            return False, None
    
    # Then check for malicious patterns
    for pattern in SQL_INJECTION_PATTERNS:
        match = re.search(pattern, query)
        if match:
            return True, match.group(0)
    
    # If no malicious pattern found
    return False, None

# Ensemble prediction function
def predict_with_ensemble(query, components):
    """
    Uses an ensemble of models to predict if the query is malicious.
    Returns predictions from individual models and ensemble vote.
    """
    # Get individual model predictions
    
    # Random Forest prediction
    query_tfidf = preprocess_query_for_traditional_ml(query, components['tfidf_vectorizer'])
    rf_pred = int(components['rf_model'].predict(query_tfidf)[0])
    
    # SVM prediction
    svm_pred = int(components['svm_model'].predict(query_tfidf)[0])
    
    # CNN prediction
    query_padded = preprocess_query_for_deep_learning(query, components['tokenizer'], components['max_sequence_length'])
    cnn_probability = components['cnn_model'].predict(query_padded)[0][0]
    cnn_pred = int(cnn_probability > 0.5)
    
    # LSTM prediction
    lstm_probability = components['lstm_model'].predict(query_padded)[0][0]
    lstm_pred = int(lstm_probability > 0.5)
    
    # Majority voting
    votes = [rf_pred, svm_pred, cnn_pred, lstm_pred]
    ensemble_pred = np.bincount(votes).argmax()
    
    return {
        'rf': rf_pred,
        'svm': svm_pred,
        'cnn': {'prediction': cnn_pred, 'probability': float(cnn_probability)},
        'lstm': {'prediction': lstm_pred, 'probability': float(lstm_probability)},
        'ensemble': int(ensemble_pred),
        'vote_count': {0: list(votes).count(0), 1: list(votes).count(1)}
    }

# Initialize session state for UI flow control
if 'analysis_stage' not in st.session_state:
    st.session_state.analysis_stage = 0  # 0: not started, 1: regex done, 2: ensemble done

if 'regex_result' not in st.session_state:
    st.session_state.regex_result = None
    
if 'ensemble_result' not in st.session_state:
    st.session_state.ensemble_result = None

# App title and description
st.title("🛡️ SQL Injection Detection")
st.markdown("""
This application uses a multi-layered approach to detect potentially malicious SQL queries:
1. **Rule-based detection** using improved regex patterns
2. **Ensemble learning** with majority voting from 4 models:
   - Random Forest
   - Support Vector Machine
   - Convolutional Neural Network
   - Long Short-Term Memory Network

Enter a query below or select from the examples to begin analysis.
""")

# Display warning if models couldn't be loaded
if model_loading_error:
    st.warning(f"⚠️ Some models could not be loaded. The application will only use rule-based detection. Error: {model_loading_error}")

# Example queries in a dropdown
st.subheader("Select an Example or Enter Your Own Query")

example_categories = {
    "Benign SQL Queries": [
    "SELECT * FROM users WHERE username='admin'",
    "SELECT id, name, price FROM products WHERE category_id=5",
    "SELECT COUNT(*) FROM orders WHERE date > '2023-01-01'",
    "INSERT INTO logs (user_id, action) VALUES (42, 'login')",
    "UPDATE customers SET last_login='2023-06-15' WHERE id=101",
    "DELETE FROM sessions WHERE last_activity < '2023-01-01'",
    "SELECT email FROM subscribers WHERE active=1",
    "INSERT INTO feedback (user_id, message) VALUES (87, 'Great service!')",
    "UPDATE inventory SET stock = stock - 1 WHERE product_id = 300",
    "SELECT name FROM employees WHERE department = 'Sales'",
    "SELECT AVG(rating) FROM reviews WHERE product_id = 55",
    "INSERT INTO audit_log (timestamp, event) VALUES (CURRENT_TIMESTAMP, 'update')",
    "SELECT * FROM appointments WHERE doctor_id = 10 AND status = 'confirmed'",
    "UPDATE settings SET value='dark' WHERE key='theme'",
    "SELECT DISTINCT city FROM customers WHERE country='USA'",
    "DELETE FROM cart_items WHERE user_id=12 AND product_id=78",
    "SELECT MAX(salary) FROM employees WHERE role='manager'",
    "INSERT INTO payments (user_id, amount, method) VALUES (33, 99.99, 'credit')",
    "UPDATE products SET price = price * 1.1 WHERE category_id = 7",
    "SELECT * FROM messages WHERE sender_id = 5 AND is_read = 0"
    ],
    "Malicious SQL Queries": [
    "' OR 1=1 --",
    "admin'; DROP TABLE users; --",
    "SELECT * FROM users WHERE username='' UNION SELECT username,password FROM admin_users --",
    "'; WAITFOR DELAY '0:0:10' --",
    "admin' OR '1'='1",
    "' OR 'a'='a",
    "' OR 1=1#",
    "' OR 1=1/*",
    "admin'--",
    "'; EXEC xp_cmdshell('dir'); --",
    "' OR EXISTS(SELECT * FROM users WHERE username = 'admin') --",
    "1; DROP TABLE sessions --",
    "'; SHUTDOWN --",
    "' OR SLEEP(5) --",
    "' AND 1=(SELECT COUNT(*) FROM users) --",
    "admin' AND SUBSTRING(password, 1, 1) = 'a' --",
    "' UNION ALL SELECT NULL,NULL,NULL --",
    "0' OR 1=1 ORDER BY 1 --",
    "1' AND (SELECT COUNT(*) FROM users) > 0 --",
    "' OR (SELECT ASCII(SUBSTRING(password,1,1)) FROM users WHERE username='admin') > 64 --"
    ]
}

# First create category selection
category = st.selectbox(
    "Choose query category:",
    options=list(example_categories.keys()),
    key="category"
)

# Then show examples from selected category
example = st.selectbox(
    "Select an example:",
    options=example_categories[category],
    key="example"
)

# Allow user to use the selected example or enter their own
query_source = st.radio(
    "Query source:",
    ["Use selected example", "Enter my own query"],
    key="query_source"
)

if query_source == "Enter my own query":
    query = st.text_area(
        "Enter SQL Query:",
        height=100,
        placeholder="Type your SQL query here..."
    )
else:
    query = example
    st.code(query, language="sql")

# Analysis process
if st.button("Start Analysis") and query:
    # Reset analysis state
    st.session_state.analysis_stage = 1
    
    # Step 1: Rule-based detection
    with st.spinner("Running rule-based detection..."):
        time.sleep(0.5)  # Simulate processing time
        is_malicious, matched_pattern = detect_sql_injection_with_regex(query)
        st.session_state.regex_result = (is_malicious, matched_pattern)

# If we have completed the regex analysis
if st.session_state.analysis_stage >= 1 and st.session_state.regex_result is not None:
    is_malicious, matched_pattern = st.session_state.regex_result
    st.subheader("Step 1: Rule-Based Detection")
    
    if is_malicious:
        st.error("🚨 SQL Injection Detected (Rule-Based)!")
        st.warning(f"Matched pattern: `{matched_pattern}`")
        
        # Show details in expander
        with st.expander("Rule-Based Detection Details"):
            st.markdown("""
            **What was detected:**
            - The query matched one or more known SQL injection patterns
            - This type of pattern is commonly used in SQL injection attacks
            - Review the query for security implications
            """)
            
            st.markdown("**Common SQL injection techniques detected:**")
            st.markdown("""
            - Comment sequences (`--`) after quotes
            - Always true conditions (`OR 1=1`)
            - Union-based injections
            - SQL command injections
            """)
    else:
        st.success("✅ No SQL injection patterns detected using rules")
        
        with st.expander("Rule-Based Detection Details"):
            st.markdown("""
            **Analysis Details:**
            - The query did not match any known SQL injection patterns
            - The structure appears to be standard SQL syntax
            - No suspicious patterns were identified
            """)
    
    # Ask if user wants to proceed with ensemble detection
    proceed = st.radio(
        "Would you like to proceed with ensemble model detection?",
        ["Yes", "No"],
        index=0,  # Default to Yes
        key="proceed"
    )
    
    # Check if models are loaded before allowing ensemble analysis
    if proceed == "Yes" and not model_loading_error:
        if st.button("Run Ensemble Analysis"):
            st.session_state.analysis_stage = 2
            with st.spinner("Running ensemble models..."):
                time.sleep(1)  # Simulate processing time
                ensemble_results = predict_with_ensemble(query, components)
                st.session_state.ensemble_result = ensemble_results
    elif proceed == "Yes" and model_loading_error:
        st.error("Cannot run ensemble analysis because models failed to load.")

# If we have completed the ensemble analysis    
if st.session_state.analysis_stage >= 2 and st.session_state.ensemble_result is not None:
    results = st.session_state.ensemble_result
    
    st.subheader("Step 2: Ensemble Model Detection")
    
    # Create a visual representation of voting
    vote_benign = results['vote_count'][0]  
    vote_malicious = results['vote_count'][1]
    
    st.markdown(f"### Model Votes")
    
    # Create columns for the voting visualization
    col1, col2 = st.columns(2)
    
    with col1:
        st.metric("Safe Votes", vote_benign)
        
    with col2:
        st.metric("Malicious Votes", vote_malicious)
    
    # Create a progress bar to visualize the voting ratio
    vote_ratio = vote_malicious / (vote_benign + vote_malicious)
    st.progress(vote_ratio, text=f"Malicious vote ratio: {vote_ratio*100:.0f}%")
    
    # Display individual model results
    st.markdown("### Individual Model Results")
    
    model_cols = st.columns(4)
    
    with model_cols[0]:
        st.markdown("**Random Forest**")
        if results['rf'] == 1:
            st.error("⚠️ Malicious")
        else:
            st.success("✅ Safe")
    
    with model_cols[1]:
        st.markdown("**SVM**")
        if results['svm'] == 1:
            st.error("⚠️ Malicious")
        else:
            st.success("✅ Safe")
    
    with model_cols[2]:
        st.markdown("**CNN**")
        cnn_prob = results['cnn']['probability'] * 100
        if results['cnn']['prediction'] == 1:
            st.error(f"⚠️ Malicious ({cnn_prob:.1f}%)")
        else:
            st.success(f"✅ Safe ({100-cnn_prob:.1f}%)")
    
    with model_cols[3]:
        st.markdown("**LSTM**")
        lstm_prob = results['lstm']['probability'] * 100
        if results['lstm']['prediction'] == 1:
            st.error(f"⚠️ Malicious ({lstm_prob:.1f}%)")
        else:
            st.success(f"✅ Safe ({100-lstm_prob:.1f}%)")
    
    # Final ensemble verdict
    st.markdown("### Ensemble Verdict")
    if results['ensemble'] == 1:
        st.error("🚨 SQL Injection Detected by Majority Vote!")
    else:
        st.success("✅ Query deemed safe by majority vote")
    
    # Explanation in expander
    with st.expander("Ensemble Detection Details"):
        st.markdown("""
        **How ensemble voting works:**
        - Each model casts a vote (0 for safe, 1 for malicious)
        - The final decision is based on majority vote
        - This approach combines the strengths of different model architectures
        - More robust than any single model alone
        """)
        
        if results['ensemble'] == 1:
            st.markdown(f"""
            **Why was this flagged:**
            - {vote_malicious} out of 4 models identified this query as potentially malicious
            - The majority vote indicates suspicious patterns
            - This query should be carefully reviewed before execution
            """)
        else:
            st.markdown(f"""
            **Why was this considered safe:**
            - {vote_benign} out of 4 models identified this query as likely safe
            - The majority vote indicates standard SQL patterns
            - No significant red flags were detected in the ensemble
            """)
    
    # Final verdict combining both approaches
    st.subheader("Final Analysis")
    
    is_malicious_regex, _ = st.session_state.regex_result
    is_malicious_ensemble = results['ensemble'] == 1
    
    if is_malicious_regex or is_malicious_ensemble:
        st.error("⚠️ This query appears to contain SQL injection patterns. Review carefully before executing.")
    else:
        st.success("✅ This query appears safe based on both rule-based and ensemble detection.")
            
    st.info("ℹ️ Remember: Always use parameterized queries and proper input validation in production systems.")
    
    # Reset button
    if st.button("Analyze Another Query"):
        st.session_state.analysis_stage = 0
        st.session_state.regex_result = None
        st.session_state.ensemble_result = None
        st.experimental_rerun()

# Sidebar with additional info
with st.sidebar:
    st.header("About This App")
    st.markdown("""
    ### Multi-Layer Detection Process
    
    1. **Rule-Based Detection**
       - Fast, pattern-matching approach
       - Uses improved regex to identify SQL injection patterns
       - Reduces false positives with safe pattern recognition
    
    2. **Ensemble Detection**
       - Combines 4 different machine learning models:
         - Random Forest
         - Support Vector Machine (SVM)
         - Convolutional Neural Network (CNN)
         - Long Short-Term Memory Network (LSTM)
       - Final decision by majority voting
    """)
    
    st.markdown("### Machine Learning Architecture")
    st.code("""
    # Traditional ML
    - Random Forest (n_estimators=100)
    - SVM (kernel='linear')
    
    # CNN Architecture
    Sequential([
        Embedding(input_dim=10000, output_dim=128),
        Conv1D(filters=64, kernel_size=3, activation='relu'),
        MaxPooling1D(pool_size=2),
        Dropout(0.5),
        Conv1D(filters=128, kernel_size=3, activation='relu'),
        MaxPooling1D(pool_size=2),
        Flatten(),
        Dense(64, activation='relu'),
        Dropout(0.5),
        Dense(1, activation='sigmoid')
    ])
    
    # LSTM Architecture
    Sequential([
        Embedding(input_dim=10000, output_dim=128),
        Bidirectional(LSTM(64, return_sequences=True)),
        Dropout(0.5),
        Bidirectional(LSTM(32)),
        Dropout(0.5),
        Dense(32, activation='relu'),
        Dense(1, activation='sigmoid')
    ])
    """)
    
    st.markdown("### How It Works")
    st.markdown("""
    1. **Step 1:** Rule-based patterns scan for known SQL injection techniques
    2. **Step 2:** Ensemble of 4 models evaluates the query structure
    3. **Final Analysis:** Combined verdict from both approaches
    """)
    
    st.markdown("---")
    st.warning("**Note:** This is a demonstration tool, not a replacement for proper security measures.")

# Footer
st.markdown("---")
st.markdown("""
<style>
.footer {
    position: fixed;
    left: 0;
    bottom: 0;
    width: 100%;
    background-color: white;
    color: black;
    text-align: center;
    padding: 10px;
    border-top: 1px solid #e5e5e5;
}
</style>
<div class="footer">
<p>Developed with ❤️ using Streamlit | SQL Injection Detection System</p>
</div>
""", unsafe_allow_html=True)