Spaces:
Sleeping
Sleeping
File size: 15,157 Bytes
55bcd2b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 | import numpy as np
import torch
from PIL import Image, ImageFilter
from app.core.config import IMAGE_ENSEMBLE, IMAGE_FAST_ENSEMBLE
from app.core.device import DEVICE
from app.models.loader import load_image_model
# ββ Model inference ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def _infer_fake_score(proc, model, img: Image.Image) -> float:
"""
Stable inference: average over 3 passes to reduce variance.
Dynamically resolves fake/real indices from id2label, no hardcoded assumptions.
Returns a score 0β1 (1 = synthetic/fake).
"""
inputs = proc(images=img, return_tensors="pt").to(DEVICE)
with torch.no_grad():
logits_list = [model(**inputs).logits for _ in range(3)]
logits_mean = torch.stack(logits_list).mean(dim=0)
probs = torch.nn.functional.softmax(logits_mean, dim=-1)[0].cpu().numpy()
id2label = {int(k): v.lower() for k, v in model.config.id2label.items()}
fake_kw = ["fake", "ai", "artificial", "synthetic", "generated", "deepfake"]
real_kw = ["real", "human", "authentic", "genuine"]
fake_indices = [i for i, lbl in id2label.items() if any(w in lbl for w in fake_kw)]
real_indices = [i for i, lbl in id2label.items() if any(w in lbl for w in real_kw)]
if not fake_indices and not real_indices:
return float(probs[1]) if len(probs) >= 2 else 0.5
fake_score = float(np.sum([probs[i] for i in fake_indices])) if fake_indices else 0.0
real_score = float(np.sum([probs[i] for i in real_indices])) if real_indices else 0.0
total = fake_score + real_score
return fake_score / total if total > 1e-9 else 0.5
def _run_ensemble(img: Image.Image, ensemble: list) -> dict:
"""Run all models in the ensemble and return weighted score + per-model details."""
results = {}
weighted_sum = 0.0
total_weight = 0.0
for cfg in ensemble:
loaded = load_image_model(cfg)
if loaded is None:
print(f" {cfg['key']} skipped (load failed)")
continue
proc, model = loaded
try:
score = _infer_fake_score(proc, model, img)
results[cfg["key"]] = {"score": round(score, 4), "weight": cfg["weight"], "desc": cfg["desc"]}
weighted_sum += score * cfg["weight"]
total_weight += cfg["weight"]
print(f" [{cfg['key']}] fake={score:.4f} Γ {cfg['weight']}")
except Exception as e:
print(f" [{cfg['key']}] error: {e}")
ensemble_score = weighted_sum / total_weight if total_weight > 0 else 0.5
return {"models": results, "ensemble_score": round(ensemble_score, 4)}
# ββ Forensic layers ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def _analyze_exif(image_bytes: bytes) -> dict:
result = {"score": 0.50, "exif_absent": False, "has_camera_info": False,
"suspicious_software": False, "ai_source": None, "details": []}
try:
import piexif
exif_data = piexif.load(image_bytes)
has_content = any(len(exif_data.get(b, {})) > 0 for b in ["0th", "Exif", "GPS", "1st"])
if not has_content:
result["exif_absent"] = True
result["details"].append("EXIF absent")
return result
zeroth = exif_data.get("0th", {})
exif_ifd = exif_data.get("Exif", {})
gps_ifd = exif_data.get("GPS", {})
sw = zeroth.get(piexif.ImageIFD.Software, b"").decode("utf-8", errors="ignore").lower()
desc = zeroth.get(piexif.ImageIFD.ImageDescription, b"").decode("utf-8", errors="ignore").lower()
artist = zeroth.get(piexif.ImageIFD.Artist, b"").decode("utf-8", errors="ignore").lower()
combined = sw + " " + desc + " " + artist
ai_sources = {
"stable diffusion": "Stable Diffusion", "midjourney": "Midjourney",
"dall-e": "DALL-E", "dallΒ·e": "DALL-E", "comfyui": "ComfyUI/SD",
"automatic1111": "Automatic1111/SD", "generative": "IA GΓ©nΓ©rative",
"diffusion": "Modèle Diffusion", "novelai": "NovelAI",
"firefly": "Adobe Firefly", "imagen": "Google Imagen",
"gemini": "Google Gemini", "flux": "Flux (BFL)",
"ideogram": "Ideogram", "leonardo": "Leonardo.ai",
"adobe ai": "Adobe AI", "ai generated": "IA GΓ©nΓ©rique",
"synthid": "Google SynthID",
}
for kw, source in ai_sources.items():
if kw in combined:
result["suspicious_software"] = True
result["ai_source"] = source
result["score"] = 0.97
result["details"].append(f"Source IA dΓ©tectΓ©e: {source}")
return result
make = zeroth.get(piexif.ImageIFD.Make, b"")
cam = zeroth.get(piexif.ImageIFD.Model, b"")
iso = exif_ifd.get(piexif.ExifIFD.ISOSpeedRatings)
shut = exif_ifd.get(piexif.ExifIFD.ExposureTime)
gps = bool(gps_ifd and len(gps_ifd) > 2)
if make or cam:
result["has_camera_info"] = True
result["details"].append(
f"Appareil: {make.decode('utf-8', errors='ignore')} {cam.decode('utf-8', errors='ignore')}".strip()
)
if gps:
result["details"].append("GPS prΓ©sent")
if result["has_camera_info"] and gps and iso and shut:
result["score"] = 0.05
elif result["has_camera_info"] and (iso or shut):
result["score"] = 0.12
elif result["has_camera_info"]:
result["score"] = 0.28
else:
result["score"] = 0.55
except Exception as e:
result["exif_absent"] = True
result["details"].append(f"Erreur EXIF: {str(e)[:60]}")
return result
def _analyze_fft(img: Image.Image, fc: float = 0.0) -> dict:
result = {"score": 0.50, "details": []}
try:
gray = np.array(img.convert("L")).astype(np.float32)
mag = np.log1p(np.abs(np.fft.fftshift(np.fft.fft2(gray))))
h, w = mag.shape
cy, cx = h // 2, w // 2
Y, X = np.ogrid[:h, :w]
dist = np.sqrt((X - cx) ** 2 + (Y - cy) ** 2)
rl, rm = min(h, w) // 8, min(h, w) // 4
le = np.mean(mag[dist <= rl])
he = np.mean(mag[(dist > rl) & (dist <= rm)])
fr = he / (le + 1e-9)
tl = 0.18 if fc > 0.45 else 0.25
th = 0.85 if fc > 0.45 else 0.72
ss = 0.70 if fr < tl else (0.55 if fr > th else 0.20)
result["details"].append(f"Ratio freq. {fr:.3f}" + (" β sur-lissage IA" if fr < tl else " β"))
pr = np.sum((mag * (dist > 5)) > (np.mean(mag) + 5 * np.std(mag))) / (h * w)
ps = 0.85 if pr > 0.003 else (0.50 if pr > 0.001 else 0.15)
result["details"].append(f"Pics GAN: {pr:.4f}" + (" β οΈ" if pr > 0.003 else " β"))
result["score"] = float(0.55 * ss + 0.45 * ps)
except Exception as e:
result["details"].append(f"Erreur FFT: {str(e)[:60]}")
return result
def _analyze_texture(img: Image.Image, fc: float = 0.0) -> dict:
result = {"score": 0.50, "details": []}
try:
arr = np.array(img).astype(np.float32)
gray = np.array(img.convert("L")).astype(np.float32)
lap = np.array(img.convert("L").filter(ImageFilter.FIND_EDGES)).astype(np.float32)
nl = float(np.std(lap))
if arr.shape[2] >= 3:
r, g, b = arr[:, :, 0], arr[:, :, 1], arr[:, :, 2]
if float(np.mean(np.abs(r - g) < 1)) > 0.98 and float(np.mean(np.abs(g - b) < 1)) > 0.98:
result["score"] = 0.85
result["details"].append("Canaux RGB identiques β image IA synthΓ©tique")
return result
ts, tm = (5.0, 14.0) if fc > 0.45 else (8.0, 20.0)
ns = 0.75 if nl > 20.0 else (0.72 if nl < ts else (0.42 if nl < tm else 0.15))
result["details"].append(f"Bruit: {nl:.1f}")
h, w, bl = gray.shape[0], gray.shape[1], 32
stds = [np.std(gray[y:y + bl, x:x + bl]) for y in range(0, h - bl, bl) for x in range(0, w - bl, bl)]
u = np.std(stds) / (np.mean(stds) + 1e-9) if stds else 0.5
ul, uh = (0.20, 0.50) if fc > 0.45 else (0.30, 0.60)
us = 0.72 if u < ul else (0.38 if u < uh else 0.15)
result["details"].append(f"UniformitΓ©: {u:.3f}")
bg_ratio = float(np.mean(gray > 200))
border_std = float(np.std(gray[:h // 8, :]))
if bg_ratio > 0.50 and border_std < 6.0:
studio_score = 0.88
elif bg_ratio > 0.50 and border_std < 15.0:
studio_score = 0.82
elif bg_ratio > 0.35 and border_std < 25.0:
studio_score = 0.55
else:
studio_score = 0.10
result["details"].append(f"Fond: {bg_ratio:.0%}")
result["score"] = float(0.35 * ns + 0.25 * us + 0.40 * studio_score)
except Exception as e:
result["details"].append(f"Erreur texture: {str(e)[:60]}")
return result
def _analyze_color(img: Image.Image) -> dict:
result = {"score": 0.50, "details": []}
try:
arr = np.array(img.convert("RGB")).astype(np.float32)
r, g, b = arr[:, :, 0].flatten(), arr[:, :, 1].flatten(), arr[:, :, 2].flatten()
def channel_entropy(ch):
hist, _ = np.histogram(ch, bins=64, range=(0, 255), density=True)
hist = hist[hist > 0]
return float(-np.sum(hist * np.log2(hist + 1e-9)))
er, eg, eb = channel_entropy(r), channel_entropy(g), channel_entropy(b)
mean_entropy = (er + eg + eb) / 3.0
entropy_std = float(np.std([er, eg, eb]))
if mean_entropy > 5.2 and entropy_std < 0.15:
ent_score = 0.72
elif mean_entropy > 4.8 and entropy_std < 0.25:
ent_score = 0.45
else:
ent_score = 0.20
result["details"].append(f"Entropie couleur: {mean_entropy:.2f}")
lum = 0.299 * r + 0.587 * g + 0.114 * b
extreme_ratio = float(np.mean((lum < 8) | (lum > 247)))
ext_score = 0.65 if extreme_ratio < 0.005 else (0.35 if extreme_ratio < 0.02 else 0.15)
result["details"].append(f"Pixels extrΓͺmes: {extreme_ratio:.4f}")
result["score"] = float(0.60 * ent_score + 0.40 * ext_score)
except Exception as e:
result["details"].append(f"Erreur palette: {str(e)[:60]}")
return result
# ββ Fusion βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def _fuse(ensemble_score: float, exif_r: dict, fft_r: dict, tex_r: dict, color_r: dict) -> dict:
exif_absent = exif_r.get("exif_absent", False)
if exif_r.get("suspicious_software"):
profile = "EXIF_IA_DETECTE"
w = {"ensemble": 0.20, "exif": 0.60, "fft": 0.12, "texture": 0.05, "color": 0.03}
elif not exif_absent and exif_r["has_camera_info"] and exif_r["score"] < 0.20:
profile = "EXIF_FIABLE"
w = {"ensemble": 0.45, "exif": 0.32, "fft": 0.12, "texture": 0.07, "color": 0.04}
elif exif_absent:
profile = "EXIF_ABSENT"
w = {"ensemble": 0.52, "exif": 0.00, "fft": 0.24, "texture": 0.14, "color": 0.10}
else:
profile = "STANDARD"
w = {"ensemble": 0.48, "exif": 0.22, "fft": 0.16, "texture": 0.09, "color": 0.05}
scores = {
"ensemble": ensemble_score,
"exif": exif_r["score"],
"fft": fft_r["score"],
"texture": tex_r["score"],
"color": color_r["score"],
}
raw = sum(w[k] * scores[k] for k in w)
# Anti-false-positive guardrails
if ensemble_score < 0.35 and fft_r["score"] < 0.38:
raw = min(raw, 0.46)
if not exif_absent and exif_r["has_camera_info"] and exif_r["score"] < 0.15:
raw = min(raw, 0.82)
if exif_r.get("suspicious_software") and raw < 0.85:
raw = max(raw, 0.90)
# High-confidence ensemble override: modern diffusion models evade forensic layers;
# when all ML models agree strongly, trust them over FFT/texture/color heuristics.
if ensemble_score >= 0.80 and not exif_r.get("has_camera_info"):
raw = max(raw, ensemble_score * 0.90)
if ensemble_score <= 0.20:
raw = min(raw, ensemble_score * 1.10 + 0.05)
return {
"fake_prob": round(raw, 4),
"real_prob": round(1.0 - raw, 4),
"layer_scores": {k: round(v, 4) for k, v in scores.items()},
"weights_used": {k: round(v, 2) for k, v in w.items()},
"fusion_profile": profile,
"ai_source": exif_r.get("ai_source"),
}
# ββ Verdict ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def _verdict(fake_prob: float, details: dict) -> dict:
if fake_prob > 0.65:
verdict = "DEEPFAKE"
confidence = "haute" if fake_prob > 0.85 else "moyenne"
reason = "Artefacts de synthèse détectés."
elif fake_prob < 0.35:
verdict = "AUTHENTIQUE"
confidence = "haute" if fake_prob < 0.15 else "moyenne"
reason = "Aucun artefact de synthèse détecté."
else:
verdict = "INDΓTERMINΓ"
confidence = "faible"
reason = "Signal ambigu, analyse non concluante."
if details.get("ai_source"):
reason = f"Source IA identifiΓ©e dans les mΓ©tadonnΓ©es: {details['ai_source']}."
return {"verdict": verdict, "confidence": confidence, "reason": reason}
# ββ Public API βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def run(img: Image.Image, image_bytes: bytes) -> dict:
"""Full analysis: 3-model ensemble + forensic layers."""
ensemble_result = _run_ensemble(img, IMAGE_ENSEMBLE)
exif_r = _analyze_exif(image_bytes)
fft_r = _analyze_fft(img)
tex_r = _analyze_texture(img)
color_r = _analyze_color(img)
fusion = _fuse(ensemble_result["ensemble_score"], exif_r, fft_r, tex_r, color_r)
verdict = _verdict(fusion["fake_prob"], fusion)
return {**verdict, **fusion, "models": ensemble_result["models"]}
def run_fast(img: Image.Image, image_bytes: bytes) -> dict:
"""Fast analysis: 2-model ensemble + EXIF only."""
ensemble_result = _run_ensemble(img, IMAGE_FAST_ENSEMBLE)
exif_r = _analyze_exif(image_bytes)
fft_r = {"score": 0.50, "details": []}
tex_r = {"score": 0.50, "details": []}
color_r = {"score": 0.50, "details": []}
fusion = _fuse(ensemble_result["ensemble_score"], exif_r, fft_r, tex_r, color_r)
verdict = _verdict(fusion["fake_prob"], fusion)
return {**verdict, **fusion, "models": ensemble_result["models"]}
|