app.py

import gradio as gr
import tensorflow as tf
import numpy as np
import tempfile
import cv2
import os

# Load model
model = tf.keras.models.load_model("deepfake_model.h5")

# Frame extractor (extracts 9 evenly spaced frames)
def extract_frames(video_path, num_frames=9):
    cap = cv2.VideoCapture(video_path)
    frames = []
    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
    interval = max(total_frames // num_frames, 1)

    count = 0
    while len(frames) < num_frames and cap.isOpened():
        ret, frame = cap.read()
        if not ret:
            break
        if count % interval == 0:
            frame = cv2.resize(frame, (224, 224))
            frames.append(frame)
        count += 1

    cap.release()

    while len(frames) < num_frames:
        frames.append(np.zeros((224, 224, 3), dtype=np.uint8))

    return np.array(frames)

# Prediction function
def predict_video(video):
    # Save uploaded video to temp file
    with tempfile.NamedTemporaryFile(delete=False, suffix=".mp4") as temp_video_file:
        with open(video, "rb") as f:
            temp_video_file.write(f.read())
        temp_video_path = temp_video_file.name

    # Extract frames
    frames = extract_frames(temp_video_path)

    # Normalize and expand dims for batch input
    frames = frames / 255.0
    input_data = np.expand_dims(frames, axis=0)

    # Predict
    prediction = model.predict(input_data)[0][0]

    # Clean up
    os.remove(temp_video_path)

    # Return result
    result = "FAKE" if prediction > 0.5 else "REAL"
    confidence = f"{prediction:.2f}" if result == "FAKE" else f"{1 - prediction:.2f}"
    return f"Prediction: {result} (Confidence: {confidence})"

# Gradio UI
iface = gr.Interface(
    fn=predict_video,
    inputs=gr.Video(),
    outputs="text",
    title="Deepfake Detection",
    description="Upload a video to detect whether it's real or fake."
)

iface.launch()