app.py

import io
from typing import List, Tuple, Dict, Any
from PIL import Image
import numpy as np
import torch
import gradio as gr

# Face detector
from facenet_pytorch import MTCNN

# HF image classifier
from transformers import AutoImageProcessor, AutoModelForImageClassification

# ========= Config =========
# You can change the model below to another public model on Hugging Face
# Example: prithivMLmods/Deep-Fake-Detector-v2-Model (binary: Deepfake vs Realism)
MODEL_ID = "prithivMLmods/Deep-Fake-Detector-v2-Model"
DEVICE = "cpu"  # Use "cuda" if GPU is available
MAX_SIDE = 640  # Resize to keep the longest side ≤ 640px for efficiency
# =========================

# ---- Utilities ----
def resize_keep_ratio(img: Image.Image, max_side: int = MAX_SIDE) -> Image.Image:
    """Resize the image while keeping aspect ratio and limit max side length."""
    w, h = img.size
    m = max(w, h)
    if m <= max_side:
        return img
    scale = max_side / float(m)
    return img.resize((int(w * scale), int(h * scale)), Image.LANCZOS)

def canonical_label(label: str) -> str:
    """Map model-specific labels to canonical 'fake' or 'real' categories."""
    l = label.lower()
    if any(k in l for k in ["fake", "ai", "synthetic", "deepfake"]):
        return "fake"
    if any(k in l for k in ["real", "authentic", "genuine"]):
        return "real"
    # Default fallback if label doesn't match known keywords
    return label

def rank_probs(id2label: Dict[int, str], probs: List[float]) -> List[Tuple[str, float]]:
    """Return sorted list of (label, probability) pairs."""
    pairs = [(id2label[i], float(probs[i])) for i in range(len(probs))]
    return sorted(pairs, key=lambda x: x[1], reverse=True)

# ---- Load models (once) ----
mtcnn = MTCNN(keep_all=True, device=DEVICE)
processor = AutoImageProcessor.from_pretrained(MODEL_ID)
clf = AutoModelForImageClassification.from_pretrained(MODEL_ID).to(DEVICE)
id2label = clf.config.id2label

# ---- Core inference ----
@torch.no_grad()
def classify_pil(img: Image.Image) -> Dict[str, Any]:
    """Run classification on a single PIL image and return ranked probabilities."""
    inputs = processor(images=img, return_tensors="pt")
    inputs = {k: v.to(DEVICE) for k, v in inputs.items()}
    logits = clf(**inputs).logits
    probs = torch.softmax(logits, dim=-1).squeeze().tolist()
    ranked = rank_probs(id2label, probs)

    # Extract approximate fake / real probabilities based on label keywords
    fake_p, real_p = None, None
    for lbl, p in ranked:
        cat = canonical_label(lbl)
        if cat == "fake" and fake_p is None:
            fake_p = p
        if cat == "real" and real_p is None:
            real_p = p
    return {
        "top": ranked[:3],
        "fake_prob": fake_p,
        "real_prob": real_p
    }

def analyze(img: Image.Image) -> Dict[str, Any]:
    """Main analysis pipeline: detect faces, classify each face or full image."""
    if img is None:
        return {"error": "No image uploaded."}

    img = img.convert("RGB")
    img = resize_keep_ratio(img, MAX_SIDE)

    # 1) Face detection
    boxes, _ = mtcnn.detect(img)
    crops = []
    if boxes is not None:
        for (x1, y1, x2, y2) in boxes:
            x1 = max(0, int(x1)); y1 = max(0, int(y1))
            x2 = min(img.width, int(x2)); y2 = min(img.height, int(y2))
            if x2 > x1 and y2 > y1:
                crops.append(img.crop((x1, y1, x2, y2)))

    results = []
    if crops:
        # 2) Classify each detected face
        for idx, face in enumerate(crops, 1):
            r = classify_pil(face)
            results.append({"face": idx, **r})
    else:
        # 3) If no face is detected, classify the whole image
        r = classify_pil(img)
        results.append({"face": None, **r})

    # Aggregate: use median of fake probabilities across all faces
    fake_scores = []
    for r in results:
        if r.get("fake_prob") is not None:
            fake_scores.append(r["fake_prob"])
        else:
            # Fallback: use top-1 label keyword
            top1 = r["top"][0][0]
            fake_scores.append(1.0 if canonical_label(top1) == "fake" else 0.0)

    if fake_scores:
        overall_fake = float(np.median(fake_scores))
    else:
        overall_fake = 0.5

    label = "Likely AI/Deepfake" if overall_fake >= 0.6 else ("Uncertain" if overall_fake >= 0.4 else "Likely Real")

    return {
        "label": label,
        "overall_fake_probability": round(overall_fake, 3),
        "faces_detected": len(crops),
        "per_face_results": results
    }

# ---- Gradio UI ----
with gr.Blocks() as demo:
    gr.Markdown(
        """
        # 🕵️ FakeSpotter — Image Deepfake Detector (CPU)
        Upload an image. If a face is detected, each face is analyzed; otherwise, the whole image is classified.  
        **No EXIF is used.** Model can be swapped by editing `MODEL_ID` in the code.  
        > Classroom demo — not a forensic tool.
        """
    )
    with gr.Row():
        inp = gr.Image(type="pil", label="Upload image")
        out = gr.JSON(label="Results")
    gr.Button("Analyze").click(analyze, inputs=inp, outputs=out)

if __name__ == "__main__":
    demo.launch()