Spaces:
Runtime error
Runtime error
File size: 8,625 Bytes
d23039a fba30db d23039a fba30db d23039a 4a4a43d d23039a 4a4a43d d23039a fba30db d23039a 4a4a43d d23039a 4a4a43d d23039a | 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 | from __future__ import annotations
import io
from typing import List
import numpy as np
from loguru import logger
from PIL import Image
from schemas.common import ArtifactIndicator
def _severity_from_score(score: float) -> str:
if score >= 0.7:
return "high"
if score >= 0.4:
return "medium"
return "low"
# ---------- 1. GAN high-frequency signature (FFT) ----------
def detect_gan_hf_artifact(pil_img: Image.Image) -> ArtifactIndicator | None:
"""Compute high-frequency energy ratio on the luminance channel.
Real photos typically follow a ~1/f spectrum; many GAN outputs show
elevated HF energy or spectral peaks.
"""
try:
gray = np.asarray(pil_img.convert("L"), dtype=np.float32)
# downsample for speed
if max(gray.shape) > 512:
import cv2
scale = 512 / max(gray.shape)
gray = cv2.resize(gray, (int(gray.shape[1] * scale), int(gray.shape[0] * scale)))
fft = np.fft.fftshift(np.fft.fft2(gray))
mag = np.abs(fft)
h, w = mag.shape
cy, cx = h // 2, w // 2
y, x = np.ogrid[:h, :w]
r = np.sqrt((x - cx) ** 2 + (y - cy) ** 2)
r_max = np.sqrt(cx * cx + cy * cy)
hf_mask = r > (0.5 * r_max)
total = float(mag.sum() + 1e-9)
hf = float(mag[hf_mask].sum())
ratio = hf / total
# Passport/ID portraits often have strong fine detail from hair, fabric,
# sharpening, and JPEG ringing. Treat this as a weak forensic signal
# unless it is extreme; the classifier ensemble remains authoritative.
score = max(0.0, min(1.0, (ratio - 0.24) / 0.28))
sev = _severity_from_score(score)
return ArtifactIndicator(
type="gan_artifact",
severity=sev,
description=(
"Unnatural pixel noise detected (common in AI-generated images)" if score > 0.4
else "Pixel noise is within expected range for a natural photo"
),
confidence=float(score),
)
except Exception as e: # noqa: BLE001
logger.warning(f"GAN HF detection failed: {e}")
return None
# ---------- 2. JPEG quantization table anomaly ----------
_STANDARD_Q_SUMS = { # rough heuristic: camera JPEGs fall in these ranges
50: (1500, 4500),
75: (600, 2500),
90: (200, 1000),
95: (100, 600),
}
def detect_compression_anomaly(raw_bytes: bytes) -> ArtifactIndicator | None:
"""Inspect JPEG quantization tables. Missing tables, non-standard layouts,
or re-saved tables often indicate manipulation or re-encoding.
"""
try:
img = Image.open(io.BytesIO(raw_bytes))
if img.format != "JPEG":
return ArtifactIndicator(
type="compression",
severity="low",
description=f"Non-JPEG format ({img.format}); compression signature not available",
confidence=0.1,
)
q = getattr(img, "quantization", None)
if not q:
return ArtifactIndicator(
type="compression",
severity="low",
description="No JPEG quantization tables readable",
confidence=0.2,
)
tables = list(q.values())
sums = [int(sum(t)) for t in tables]
num_tables = len(tables)
# Heuristics: very low sum → very high quality (possibly re-saved);
# non-standard number of tables; extreme values.
suspicious = 0.0
reasons: list[str] = []
if num_tables not in (1, 2):
suspicious += 0.4
reasons.append(f"unusual table count ({num_tables})")
if any(s < 60 for s in sums):
suspicious += 0.3
reasons.append("very low quantization sums (possible re-encoding)")
if any(s > 8000 for s in sums):
suspicious += 0.2
reasons.append("very high quantization sums")
score = max(0.0, min(1.0, suspicious))
sev = _severity_from_score(score)
desc = (
"Signs of repeated image compression or manipulation detected" if reasons else "No unusual compression artifacts detected"
)
return ArtifactIndicator(
type="compression",
severity=sev,
description=desc,
confidence=float(score),
)
except Exception as e: # noqa: BLE001
logger.warning(f"Compression anomaly detection failed: {e}")
return None
# ---------- 3. Facial boundary + 4. Lighting (MediaPipe) ----------
def detect_face_based_artifacts(pil_img: Image.Image) -> List[ArtifactIndicator]:
"""If a face is detected, analyze jaw boundary variance and per-quadrant
luminance balance. Returns 0, 1, or 2 indicators.
"""
results: List[ArtifactIndicator] = []
try:
from models.model_loader import get_model_loader
detector = get_model_loader().load_face_detector()
if detector is None:
return results
rgb = np.asarray(pil_img.convert("RGB"))
h, w = rgb.shape[:2]
mp_result = detector.process(rgb)
if not mp_result.multi_face_landmarks:
return results
landmarks = mp_result.multi_face_landmarks[0].landmark
# ----- Jaw boundary jitter -----
# FaceMesh jaw/oval landmark indices (approximate face contour)
JAW_IDX = [
10, 338, 297, 332, 284, 251, 389, 356, 454, 323, 361,
288, 397, 365, 379, 378, 400, 377, 152, 148, 176, 149,
150, 136, 172, 58, 132, 93, 234, 127, 162, 21, 54, 103, 67, 109,
]
pts = np.array([(landmarks[i].x * w, landmarks[i].y * h) for i in JAW_IDX])
# Second-difference magnitude = local curvature jitter
diffs = np.diff(pts, axis=0)
seconds = np.diff(diffs, axis=0)
jitter = float(np.linalg.norm(seconds, axis=1).mean()) / max(w, h)
jitter_score = max(0.0, min(1.0, (jitter - 0.003) / 0.010))
results.append(
ArtifactIndicator(
type="facial_boundary",
severity=_severity_from_score(jitter_score),
description=(
"Inconsistent blending detected around the facial boundary" if jitter_score > 0.4
else "Facial boundary blending appears smooth and natural"
),
confidence=float(jitter_score),
)
)
# ----- Lighting inconsistency (per-quadrant luminance) -----
xs = np.array([lm.x * w for lm in landmarks])
ys = np.array([lm.y * h for lm in landmarks])
x0, x1 = int(max(0, xs.min())), int(min(w, xs.max()))
y0, y1 = int(max(0, ys.min())), int(min(h, ys.max()))
if x1 > x0 + 4 and y1 > y0 + 4:
face_crop = rgb[y0:y1, x0:x1]
gray = 0.299 * face_crop[..., 0] + 0.587 * face_crop[..., 1] + 0.114 * face_crop[..., 2]
hh, ww = gray.shape
quads = [
gray[: hh // 2, : ww // 2],
gray[: hh // 2, ww // 2 :],
gray[hh // 2 :, : ww // 2],
gray[hh // 2 :, ww // 2 :],
]
means = np.array([q.mean() for q in quads if q.size > 0])
if means.size == 4 and means.mean() > 1e-3:
imbalance = float(means.std() / means.mean())
lighting_score = max(0.0, min(1.0, (imbalance - 0.08) / 0.20))
results.append(
ArtifactIndicator(
type="lighting",
severity=_severity_from_score(lighting_score),
description=(
"Unnatural or inconsistent lighting detected across the face" if lighting_score > 0.4
else "Facial lighting appears natural and uniform"
),
confidence=float(lighting_score),
)
)
except Exception as e: # noqa: BLE001
logger.warning(f"Face-based artifact detection failed: {e}")
return results
# ---------- Orchestrator ----------
def scan_artifacts(pil_img: Image.Image, raw_bytes: bytes) -> List[ArtifactIndicator]:
indicators: List[ArtifactIndicator] = []
for fn in (
lambda: detect_gan_hf_artifact(pil_img),
lambda: detect_compression_anomaly(raw_bytes),
):
ind = fn()
if ind is not None:
indicators.append(ind)
indicators.extend(detect_face_based_artifacts(pil_img))
return indicators
|