app.py

import gradio as gr
import tensorflow as tf
import pickle
import numpy as np
from sklearn.preprocessing import LabelEncoder

# Load saved components
with open('preprocessing_params.pkl', 'rb') as f:
    preprocessing_params = pickle.load(f)
with open('fisher_information.pkl', 'rb') as f:
    fisher_information = pickle.load(f)
with open('label_encoder.pkl', 'rb') as f:
    label_encoder = pickle.load(f)
with open('url_tokenizer.pkl', 'rb') as f:
    url_tokenizer = pickle.load(f)
with open('html_tokenizer.pkl', 'rb') as f:
    html_tokenizer = pickle.load(f)

# Load the model with custom loss
@tf.keras.utils.register_keras_serializable()
class EWCLoss(tf.keras.losses.Loss):
    def __init__(self, model=None, fisher_information=None, importance=1.0, reduction='auto', name=None):
        super(EWCLoss, self).__init__(reduction=reduction, name=name)
        self.model = model
        self.fisher_information = fisher_information
        self.importance = importance
        self.prev_weights = [layer.numpy() for layer in model.trainable_weights] if model else None

    def call(self, y_true, y_pred):
        standard_loss = tf.keras.losses.binary_crossentropy(y_true, y_pred)
        ewc_loss = 0.0
        for layer, fisher_info, prev_weight in zip(self.model.trainable_weights, self.fisher_information, self.prev_weights):
            ewc_loss += tf.reduce_sum(fisher_info * tf.square(layer - prev_weight))
        return standard_loss + (self.importance / 2.0) * ewc_loss
    
    def get_config(self):
        config = super().get_config()
        config.update({
            'importance': self.importance,
            'reduction': self.reduction,
            'name': self.name,
        })
        return config
    
    @classmethod
    def from_config(cls, config):
        with open('fisher_information.pkl', 'rb') as f:
            fisher_information = pickle.load(f)
        return cls(model=None, fisher_information=fisher_information, **config)

# Load the model first without the custom loss
model = tf.keras.models.load_model('new_phishing_detection_model.keras', compile=False)

# Reconstruct the EWC loss
ewc_loss = EWCLoss(model=model, fisher_information=fisher_information, importance=1000)

# Compile the model with EWC loss and metrics
model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.0005),
              loss=ewc_loss,
              metrics=['accuracy', tf.keras.metrics.Precision(), tf.keras.metrics.Recall()])

# Function to preprocess input
def preprocess_input(input_text, tokenizer, max_length):
    sequences = tokenizer.texts_to_sequences([input_text])
    padded_sequences = tf.keras.preprocessing.sequence.pad_sequences(sequences, maxlen=max_length, padding='post', truncating='post')
    return padded_sequences

# Function to get prediction
def get_prediction(input_text, input_type):
    is_url = input_type == "URL"
    if is_url:
        input_data = preprocess_input(input_text, url_tokenizer, preprocessing_params['max_url_length'])
    else:
        input_data = preprocess_input(input_text, html_tokenizer, preprocessing_params['max_html_length'])
    
    prediction = model.predict([input_data, input_data])[0][0]
    return prediction

# Gradio UI
def phishing_detection(input_text, input_type):
    prediction = get_prediction(input_text, input_type)
    if prediction > 0.5:
        return f"Warning: This site is likely a phishing site! ({prediction:.2f})"
    else:
        return f"Safe: This site is not likely a phishing site. ({prediction:.2f})"

iface = gr.Interface(
    fn=phishing_detection,
    inputs=[
        gr.components.Textbox(lines=5, placeholder="Enter URL or HTML code"), 
        gr.components.Radio(["URL", "HTML"], type="value", label="Input Type")
    ],
    outputs=gr.components.Textbox(label="Phishing Detection Result"),
    title="Phishing Detection with Enhanced EWC Model",
    description="Check if a URL or HTML is Phishing.",
    theme="default"
)

iface.launch()