Instagram終焉へ。3～5年で全画像AI生成時代突入、真偽判定不可能に

# c2pa-pythonライブラリを使ったメタデータ検証
# pip install c2pa-python
import c2pa

def verify_image_provenance(image_path: str) -> dict:
    """
    画像のC2PAメタデータを検証する。
    生成AIが埋め込んだ来歴情報の有無を確認。
    """
    reader = c2pa.Reader.from_file(image_path)
    
    if reader is None:
        return {"status": "no_c2pa_data", "is_ai_generated": None}
    
    manifest = reader.get_active_manifest()
    assertions = manifest.assertions() if manifest else []
    
    ai_assertions = [
        a for a in assertions 
        if "ai.generated" in a.label() or "c2pa.training-mining" in a.label()
    ]
    
    return {
        "status": "verified",
        "is_ai_generated": len(ai_assertions) > 0,
        "assertions": [a.label() for a in ai_assertions],
        "issuer": manifest.claim_generator() if manifest else None
    }

# 実行例
result = verify_image_provenance("sample_instagram_photo.jpg")
print(result)
# → {'status': 'verified', 'is_ai_generated': True, 'assertions': ['c2pa.ai.generated'], ...}

# CNNDetect（Wang et al., 2020）ベースの検出例
# 事前学習済みモデルを使った二値分類
import torch
import torchvision.transforms as transforms
from PIL import Image

class AIImageDetector:
    def __init__(self, model_path: str):
        self.model = torch.load(model_path, map_location="cpu")
        self.model.eval()
        self.transform = transforms.Compose([
            transforms.Resize((224, 224)),
            transforms.ToTensor(),
            transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
        ])
    
    def predict(self, image_path: str) -> dict:
        img = Image.open(image_path).convert("RGB")
        tensor = self.transform(img).unsqueeze(0)
        
        with torch.no_grad():
            logit = self.model(tensor)
            prob = torch.sigmoid(logit).item()
        
        return {
            "ai_probability": round(prob, 4),
            "verdict": "AI生成の可能性が高い" if prob > 0.5 else "実写の可能性が高い",
            "confidence": "高" if abs(prob - 0.5) > 0.3 else "低"
        }

detector = AIImageDetector("resnet50_finetuned_detector.pth")
result = detector.predict("suspicious_photo.jpg")
print(result)
# → {'ai_probability': 0.9231, 'verdict': 'AI生成の可能性が高い', 'confidence': '高'}

# invisible-watermark ライブラリを使った透かし埋め込み
# pip install invisible-watermark
from imwatermark import WatermarkEncoder, WatermarkDecoder
import cv2
import numpy as np

def embed_watermark(image_path: str, output_path: str, wm_text: str = "AI_GENERATED") -> None:
    """
    画像にDWT-DCT-SVDベースの不可視透かしを埋め込む。
    人間の目には検知不可能だが、デコーダで復元可能。
    """
    bgr_img = cv2.imread(image_path)
    
    encoder = WatermarkEncoder()
    # 48ビットのバイナリウォーターマークとして埋め込む
    wm_bytes = wm_text.encode("utf-8")[:6].ljust(6, b"\x00")
    encoder.set_watermark("bytes", wm_bytes)
    
    watermarked = encoder.encode(bgr_img, "dwtDctSvd")
    cv2.imwrite(output_path, watermarked)
    print(f"透かし埋め込み完了: {output_path}")

def decode_watermark(image_path: str) -> str:
    """埋め込まれた透かしを復元する"""
    bgr_img = cv2.imread(image_path)
    decoder = WatermarkDecoder("bytes", 48)
    watermark = decoder.decode(bgr_img, "dwtDctSvd")
    return watermark.decode("utf-8", errors="ignore").strip("\x00")

# 使用例
embed_watermark("generated_photo.png", "watermarked_photo.png", "FLUX_GENERATED")
result = decode_watermark("watermarked_photo.png")
print(f"検出された透かし: {result}")  # → 検出された透かし: FLUX_GENERATED