# deepfake_scanner.py
import os, json, requests, subprocess
from pathlib import Path
import torch
from detectron2.engine import DefaultPredictor
from detectron2.config import get_cfg
from detectron2 import model_zoo

# 1️⃣ Configuración del modelo Detectron (pre‑entrenado en FaceForensics++)
cfg = get_cfg()
cfg.merge_from_file(model_zoo.get_config_file(
    "COCO-Detection/faster_rcnn_R_50_FPN_3x.yaml"))
cfg.MODEL.WEIGHTS = model_zoo.get_checkpoint_url(
    "COCO-Detection/faster_rcnn_R_50_FPN_3x.yaml")
cfg.MODEL.DEVICE = "cuda"
predictor = DefaultPredictor(cfg)

# 2️⃣ Función para extraer frames cada segundo
def extract_frames(video_path: Path, out_dir: Path):
    out_dir.mkdir(parents=True, exist_ok=True)
    cmd = [
        "ffmpeg", "-i", str(video_path),
        "-vf", "fps=1",               # 1 frame por segundo
        str(out_dir / "frame_%04d.jpg")
    ]
    subprocess.run(cmd, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)

# 3️⃣ Analizar cada frame con Detectron
def analyze_frames(frames_dir: Path) -> float:
    scores = []
    for img_path in sorted(frames_dir.glob("*.jpg")):
        im = torch.imread(str(img_path)).permute(2, 0, 1)  # C,H,W
        outputs = predictor(im)
        # En los modelos de forensics, un alto “score” de cara falsa >0.7 indica deepfake
        for inst in outputs["instances"]:
            if inst.scores.item() > 0.7:
                scores.append(inst.scores.item())
    return sum(scores) / len(scores) if scores else 0.0

# 4️⃣ Consulta a Deepware (API pública)
def deepware_api(video_path: Path) -> dict:
    url = "https://api.deepware.ai/v1/scan"
    files = {"file": open(video_path, "rb")}
    resp = requests.post(url, files=files)
    return resp.json()

# 5️⃣ Pipeline completo
def scan_video(video_path: str):
    video = Path(video_path)
    tmp_dir = Path("./tmp_frames") / video.stem
    extract_frames(video, tmp_dir)

    local_score = analyze_frames(tmp_dir)
    api_result = deepware_api(video)

    # Resultado consolidado
    print(f"\n🔎 Análisis de {video.name}")
    print(f"→ Score Detectron (0‑1): {local_score:.3f}")
    print(f"→ Deepware confidence: {api_result.get('confidence', 'N/A')}")
    print("→ Verdict: " + ("POSIBLE DEEPFAKE" if local_score > 0.6 or api_result.get('is_fake') else "PARECE LEGÍTIMO"))

    # Limpieza
    subprocess.run(["rm", "-rf", str(tmp_dir)])

if __name__ == "__main__":
    import sys
    if len(sys.argv) != 2:
        print("Uso: python deepfake_scanner.py <ruta_video>")
        sys.exit(1)
    scan_video(sys.argv[1])

pip install torch torchvision detectron2 opencv-python requests

python deepfake_scanner.py campaña_2026.mp4