Add DeepGuard deepfake detection app with XceptionTransfer model

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+*.keras filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,14 +1,73 @@
 ---
-title: DeepGuard
-emoji: ⚡
-colorFrom: yellow
+title: DeepGuard AI Face Authenticator
+emoji: 🛡️
+colorFrom: green
 colorTo: blue
 sdk: gradio
-sdk_version: 6.1.0
+sdk_version: 4.44.0
 app_file: app.py
 pinned: false
 license: mit
-short_description: Real / Fake? Upload a face and let AI forensics reveal truth
+tags:
+  - deepfake-detection
+  - computer-vision
+  - tensorflow
+  - mlops
+short_description: Deepfake face detection using XceptionTransfer with 88% accuracy
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# DeepGuard: AI Face Authenticator
+
+A production-grade deep learning system for detecting AI-generated (deepfake) faces, built with a complete MLOps pipeline.
+
+## Overview
+
+DeepGuard leverages transfer learning with the Xception architecture to identify synthetic faces generated by GANs (Generative Adversarial Networks). The model achieves 88% accuracy with a 95% ROC-AUC score on StyleGAN-generated content.
+
+## Model Performance
+
+| Metric | Value |
+|--------|-------|
+| Test Accuracy | 88% |
+| ROC-AUC Score | 95% |
+| Training Dataset | 140,000 images |
+| Architecture | XceptionTransfer |
+| Input Resolution | 128x128 pixels |
+
+## FFT Frequency Analysis Interpretation
+
+The Fast Fourier Transform visualization provides forensic insight into image authenticity.
+
+| Pattern | Interpretation |
+|---------|----------------|
+| Bright center spot | Normal low-frequency content (smooth areas) |
+| Radiating spokes | Edge directions in the original image |
+| Random noise distribution | Natural texture typical of real photographs |
+| Grid or cross artifacts | Potential GAN fingerprint indicating synthetic generation |
+
+Note: GAN artifacts in the frequency domain are subtle and serve as a supplementary forensic tool.
+
+## Known Limitations
+
+This model is trained on StyleGAN-generated faces. Detection accuracy may be reduced for:
+
+- Images from diffusion models (Stable Diffusion, Midjourney, DALL-E)
+- Non-face subjects or full-body photographs
+- Heavily compressed or filtered images
+
+## MLOps Pipeline
+
+| Component | Technology |
+|-----------|------------|
+| Data Versioning | DVC |
+| Experiment Tracking | MLflow + DagsHub |
+| Model Training | TensorFlow / Keras |
+| Deployment | Hugging Face Spaces |
+
+## Repository
+
+[GitHub: DeepGuard-MLOps-Pipeline](https://github.com/HarshTomar1234/DeepGuard-MLOps-Pipeline)
+
+## License
+
+MIT License

app.py

@@ -0,0 +1,135 @@
+import gradio as gr
+import tensorflow as tf
+import numpy as np
+from PIL import Image
+
+# Load model
+print("Loading DeepGuard model...")
+model = tf.keras.models.load_model("deepfake_model.keras")
+print("Model loaded successfully!")
+
+
+def generate_fft(image):
+    """Generate FFT magnitude spectrum visualization."""
+    try:
+        img_gray = image.convert('L').resize((128, 128))
+        img_array = np.array(img_gray, dtype=np.float32)
+        f_transform = np.fft.fft2(img_array)
+        f_shift = np.fft.fftshift(f_transform)
+        magnitude = np.abs(f_shift)
+        magnitude = np.log1p(magnitude)
+        magnitude = ((magnitude - magnitude.min()) / (magnitude.max() - magnitude.min()) * 255).astype(np.uint8)
+        return Image.fromarray(magnitude)
+    except Exception as e:
+        print(f"FFT Error: {e}")
+        return None
+
+
+def predict(image):
+    """Main prediction function."""
+    if image is None:
+        return None, "Please upload an image", None
+    
+    try:
+        img = image.convert('RGB').resize((128, 128))
+        img_array = np.array(img) / 255.0
+        img_array = np.expand_dims(img_array, axis=0)
+        
+        prediction = float(model.predict(img_array, verbose=0)[0][0])
+        
+        if prediction > 0.5:
+            label = "FAKE (AI Generated)"
+            confidence = prediction * 100
+        else:
+            label = "REAL (Authentic)"
+            confidence = (1 - prediction) * 100
+        
+        result_text = f"""
+## Detection Result: {label}
+
+**Confidence:** {confidence:.2f}%
+
+**Raw Model Score:** {prediction:.6f}
+
+---
+
+### Interpretation
+- Score > 0.5: Model detects GAN artifacts = FAKE
+- Score < 0.5: No GAN artifacts detected = REAL
+
+### Important Note
+This model is trained on StyleGAN-generated faces and may not accurately detect images from modern diffusion models (Stable Diffusion, Midjourney, DALL-E).
+"""
+        
+        fft_image = generate_fft(image)
+        return result_text, fft_image
+        
+    except Exception as e:
+        return f"Error: {str(e)}", None
+
+
+with gr.Blocks(
+    title="DeepGuard - AI Face Authenticator",
+    theme=gr.themes.Base(primary_hue="green", secondary_hue="blue", neutral_hue="gray")
+) as demo:
+    
+    gr.HTML("""
+    <div style="text-align: center; margin-bottom: 1rem;">
+        <h1 style="color: #00f260; font-size: 2.5rem;">DeepGuard</h1>
+        <p style="color: #888;">AI Face Authenticator - Deepfake Detection</p>
+    </div>
+    """)
+    
+    with gr.Row():
+        with gr.Column(scale=1):
+            input_image = gr.Image(label="Upload Image", type="pil", height=300)
+            analyze_btn = gr.Button("Analyze Image", variant="primary", size="lg")
+    
+    with gr.Row():
+        with gr.Column(scale=2):
+            result_output = gr.Markdown(label="Detection Result")
+        with gr.Column(scale=1):
+            fft_output = gr.Image(label="FFT Frequency Analysis", height=200)
+    
+    with gr.Accordion("How It Works", open=False):
+        gr.Markdown("""
+### XceptionTransfer Deep Learning Model
+
+- **Architecture:** Transfer learning with Xception backbone
+- **Training Data:** 140,000 real and GAN-generated faces
+- **Accuracy:** 88% on test dataset
+- **ROC-AUC:** 95%
+        """)
+    
+    with gr.Accordion("FFT Interpretation Guide", open=False):
+        gr.Markdown("""
+| Pattern | Interpretation |
+|---------|----------------|
+| Bright center spot | Normal low-frequency content |
+| Radiating spokes | Edge directions in original image |
+| Random noise distribution | Natural texture (real photos) |
+| Grid or cross artifacts | Potential GAN fingerprint |
+
+*GAN artifacts in FFT are subtle and require trained interpretation.*
+        """)
+    
+    with gr.Accordion("Model Limitations", open=False):
+        gr.Markdown("""
+This model is trained on StyleGAN-generated faces only.
+
+It may NOT accurately detect:
+- Stable Diffusion / Midjourney / DALL-E images
+- Non-face subjects
+- Heavily edited images
+        """)
+    
+    gr.HTML("""<div style="text-align: center; margin-top: 2rem; color: #666;">
+        Built with TensorFlow and MLflow | DeepGuard MLOps Pipeline
+    </div>""")
+    
+    analyze_btn.click(fn=predict, inputs=input_image, outputs=[result_output, fft_output])
+    input_image.change(fn=predict, inputs=input_image, outputs=[result_output, fft_output])
+
+
+if __name__ == "__main__":
+    demo.launch()

deepfake_model.keras

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e3b3d51181f4d1fa2678b6b822ba9e4becf67bb72231dfc693d42f1aff491dc3
+size 169750460

requirements.txt

@@ -0,0 +1,4 @@
+gradio==4.44.0
+tensorflow==2.17.0
+numpy>=1.24.0
+Pillow>=10.0.0