Spaces:
Running
Running
BitCheck Codex commited on
Commit ·
8624fca
1
Parent(s): 662bc65
feat: add lightweight image forensic analysis
Browse files
app/services/forensic_analyzer.py
ADDED
|
@@ -0,0 +1,152 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from __future__ import annotations
|
| 2 |
+
|
| 3 |
+
from pathlib import Path
|
| 4 |
+
from typing import Any
|
| 5 |
+
|
| 6 |
+
import cv2
|
| 7 |
+
import numpy as np
|
| 8 |
+
from PIL import Image, ImageOps
|
| 9 |
+
|
| 10 |
+
|
| 11 |
+
def analyze_image_forensics(path: Path, verification_id: str, output_dir: Path) -> dict[str, Any]:
|
| 12 |
+
try:
|
| 13 |
+
with Image.open(path) as img:
|
| 14 |
+
rgb = np.array(ImageOps.exif_transpose(img).convert("RGB"))
|
| 15 |
+
gray = cv2.cvtColor(rgb, cv2.COLOR_RGB2GRAY)
|
| 16 |
+
sharpness = float(cv2.Laplacian(gray, cv2.CV_64F).var())
|
| 17 |
+
|
| 18 |
+
noise_values, boxes = _block_metric(gray, _noise_metric)
|
| 19 |
+
blur_values, _ = _block_metric(gray, _blur_metric)
|
| 20 |
+
edge_values, _ = _block_metric(gray, _edge_metric)
|
| 21 |
+
brightness_values, _ = _block_metric(gray, lambda p: float(np.mean(p)))
|
| 22 |
+
contrast_values, _ = _block_metric(gray, lambda p: float(np.std(p)))
|
| 23 |
+
|
| 24 |
+
noise_risk = _inconsistency(noise_values)
|
| 25 |
+
blur_risk = _inconsistency(blur_values)
|
| 26 |
+
edge_risk = _inconsistency(edge_values)
|
| 27 |
+
compression_risk = _compression_proxy(gray)
|
| 28 |
+
brightness_contrast_risk = max(_inconsistency(brightness_values), _inconsistency(contrast_values)) * 0.7
|
| 29 |
+
|
| 30 |
+
suspicious = _suspicious_regions(boxes, noise_values, blur_values, edge_values)
|
| 31 |
+
annotated_url = _save_annotation(rgb, suspicious, verification_id, output_dir) if suspicious else None
|
| 32 |
+
risk = 0.24 * noise_risk + 0.22 * compression_risk + 0.2 * blur_risk + 0.2 * edge_risk + 0.14 * brightness_contrast_risk
|
| 33 |
+
|
| 34 |
+
flags: list[str] = []
|
| 35 |
+
if sharpness < 45:
|
| 36 |
+
flags.append("Image has low global sharpness; this may reflect blur, resizing, or compression.")
|
| 37 |
+
if noise_risk > 0.55:
|
| 38 |
+
flags.append("Possible noise inconsistency across image regions.")
|
| 39 |
+
if compression_risk > 0.55:
|
| 40 |
+
flags.append("Possible compression artifact inconsistency.")
|
| 41 |
+
if blur_risk > 0.55:
|
| 42 |
+
flags.append("Possible local blur inconsistency.")
|
| 43 |
+
if edge_risk > 0.55:
|
| 44 |
+
flags.append("Possible edge density inconsistency.")
|
| 45 |
+
if brightness_contrast_risk > 0.55:
|
| 46 |
+
flags.append("Possible brightness or contrast anomaly.")
|
| 47 |
+
|
| 48 |
+
return {
|
| 49 |
+
"checked": True,
|
| 50 |
+
"risk_score": round(float(min(1.0, risk)), 3),
|
| 51 |
+
"sharpness_score": round(sharpness, 3),
|
| 52 |
+
"noise_inconsistency_risk": round(float(noise_risk), 3),
|
| 53 |
+
"compression_risk": round(float(compression_risk), 3),
|
| 54 |
+
"blur_inconsistency_risk": round(float(blur_risk), 3),
|
| 55 |
+
"edge_inconsistency_risk": round(float(edge_risk), 3),
|
| 56 |
+
"brightness_contrast_risk": round(float(brightness_contrast_risk), 3),
|
| 57 |
+
"suspicious_regions": suspicious,
|
| 58 |
+
"annotated_image_url": annotated_url,
|
| 59 |
+
"flags": flags,
|
| 60 |
+
"disclaimer": "Forensic indicators are risk signals, not definitive proof of manipulation.",
|
| 61 |
+
}
|
| 62 |
+
except Exception as exc:
|
| 63 |
+
return {
|
| 64 |
+
"checked": True,
|
| 65 |
+
"risk_score": 0.0,
|
| 66 |
+
"sharpness_score": 0.0,
|
| 67 |
+
"noise_inconsistency_risk": 0.0,
|
| 68 |
+
"compression_risk": 0.0,
|
| 69 |
+
"blur_inconsistency_risk": 0.0,
|
| 70 |
+
"edge_inconsistency_risk": 0.0,
|
| 71 |
+
"suspicious_regions": [],
|
| 72 |
+
"annotated_image_url": None,
|
| 73 |
+
"flags": ["Forensic analysis failed."],
|
| 74 |
+
"disclaimer": "Forensic indicators are risk signals, not definitive proof of manipulation.",
|
| 75 |
+
"error": str(exc),
|
| 76 |
+
}
|
| 77 |
+
|
| 78 |
+
|
| 79 |
+
def _block_metric(gray: np.ndarray, fn: Any, grid: int = 4) -> tuple[list[float], list[dict[str, int]]]:
|
| 80 |
+
h, w = gray.shape
|
| 81 |
+
values: list[float] = []
|
| 82 |
+
boxes: list[dict[str, int]] = []
|
| 83 |
+
ys = np.linspace(0, h, grid + 1, dtype=int)
|
| 84 |
+
xs = np.linspace(0, w, grid + 1, dtype=int)
|
| 85 |
+
for row in range(grid):
|
| 86 |
+
for col in range(grid):
|
| 87 |
+
y0, y1 = int(ys[row]), int(ys[row + 1])
|
| 88 |
+
x0, x1 = int(xs[col]), int(xs[col + 1])
|
| 89 |
+
patch = gray[y0:y1, x0:x1]
|
| 90 |
+
values.append(fn(patch))
|
| 91 |
+
boxes.append({"x": x0, "y": y0, "width": x1 - x0, "height": y1 - y0})
|
| 92 |
+
return values, boxes
|
| 93 |
+
|
| 94 |
+
|
| 95 |
+
def _noise_metric(patch: np.ndarray) -> float:
|
| 96 |
+
blur = cv2.GaussianBlur(patch, (3, 3), 0)
|
| 97 |
+
return float(np.std(patch.astype("float32") - blur.astype("float32")))
|
| 98 |
+
|
| 99 |
+
|
| 100 |
+
def _blur_metric(patch: np.ndarray) -> float:
|
| 101 |
+
return float(cv2.Laplacian(patch, cv2.CV_64F).var())
|
| 102 |
+
|
| 103 |
+
|
| 104 |
+
def _edge_metric(patch: np.ndarray) -> float:
|
| 105 |
+
edges = cv2.Canny(patch, 80, 180)
|
| 106 |
+
return float(np.mean(edges > 0))
|
| 107 |
+
|
| 108 |
+
|
| 109 |
+
def _inconsistency(values: list[float]) -> float:
|
| 110 |
+
arr = np.array(values, dtype="float32")
|
| 111 |
+
if arr.size == 0:
|
| 112 |
+
return 0.0
|
| 113 |
+
med = float(np.median(arr)) + 1e-6
|
| 114 |
+
mad = float(np.median(np.abs(arr - med))) + 1e-6
|
| 115 |
+
robust_cv = mad / (abs(med) + 1e-6)
|
| 116 |
+
return float(min(1.0, robust_cv / 0.65))
|
| 117 |
+
|
| 118 |
+
|
| 119 |
+
def _compression_proxy(gray: np.ndarray) -> float:
|
| 120 |
+
h, w = gray.shape
|
| 121 |
+
h8, w8 = h - (h % 8), w - (w % 8)
|
| 122 |
+
if h8 < 16 or w8 < 16:
|
| 123 |
+
return 0.0
|
| 124 |
+
crop = gray[:h8, :w8].astype("float32")
|
| 125 |
+
vertical = np.mean(np.abs(crop[:, 7::8] - crop[:, 8::8])) if crop[:, 8::8].size else 0.0
|
| 126 |
+
horizontal = np.mean(np.abs(crop[7::8, :] - crop[8::8, :])) if crop[8::8, :].size else 0.0
|
| 127 |
+
return float(min(1.0, (vertical + horizontal) / 36.0))
|
| 128 |
+
|
| 129 |
+
|
| 130 |
+
def _suspicious_regions(boxes: list[dict[str, int]], *metrics: list[float]) -> list[dict[str, Any]]:
|
| 131 |
+
scores = np.zeros(len(boxes), dtype="float32")
|
| 132 |
+
for values in metrics:
|
| 133 |
+
arr = np.array(values, dtype="float32")
|
| 134 |
+
med = np.median(arr)
|
| 135 |
+
std = np.std(arr) + 1e-6
|
| 136 |
+
scores += np.abs(arr - med) / std
|
| 137 |
+
regions = []
|
| 138 |
+
for box, score in sorted(zip(boxes, scores), key=lambda item: float(item[1]), reverse=True)[:5]:
|
| 139 |
+
normalized = float(min(1.0, score / 6.0))
|
| 140 |
+
if normalized >= 0.45:
|
| 141 |
+
regions.append({**box, "score": round(normalized, 3), "label": "possible visual inconsistency"})
|
| 142 |
+
return regions
|
| 143 |
+
|
| 144 |
+
|
| 145 |
+
def _save_annotation(rgb: np.ndarray, regions: list[dict[str, Any]], verification_id: str, output_dir: Path) -> str:
|
| 146 |
+
annotated = rgb.copy()
|
| 147 |
+
for region in regions:
|
| 148 |
+
x, y, w, h = region["x"], region["y"], region["width"], region["height"]
|
| 149 |
+
cv2.rectangle(annotated, (x, y), (x + w, y + h), (255, 188, 40), max(2, annotated.shape[1] // 300))
|
| 150 |
+
path = output_dir / f"{verification_id}_forensics.jpg"
|
| 151 |
+
Image.fromarray(annotated).save(path, quality=92)
|
| 152 |
+
return f"/outputs/{path.name}"
|
app/services/forensic_checker.py
CHANGED
|
@@ -1,84 +1,9 @@
|
|
| 1 |
-
from __future__ import annotations
|
| 2 |
-
|
| 3 |
-
from io import BytesIO
|
| 4 |
from pathlib import Path
|
| 5 |
from typing import Any
|
| 6 |
|
| 7 |
-
import
|
| 8 |
-
|
| 9 |
-
from PIL import Image
|
| 10 |
-
|
| 11 |
-
|
| 12 |
-
def analyze_forensics(path: Path, metadata: dict[str, Any]) -> dict[str, Any]:
|
| 13 |
-
try:
|
| 14 |
-
with Image.open(path) as img:
|
| 15 |
-
original_mode = img.mode
|
| 16 |
-
fmt = img.format
|
| 17 |
-
width, height = img.size
|
| 18 |
-
rgb = np.array(img.convert("RGB"))
|
| 19 |
-
gray = cv2.cvtColor(rgb, cv2.COLOR_RGB2GRAY)
|
| 20 |
-
sharpness = float(cv2.Laplacian(gray, cv2.CV_64F).var())
|
| 21 |
-
|
| 22 |
-
compression_risk = _jpeg_recompression_delta(rgb) if fmt == "JPEG" else 0.15
|
| 23 |
-
noise_risk = _noise_inconsistency(gray)
|
| 24 |
-
metadata_risk = 0.35 if not metadata.get("metadata_found") else 0.05
|
| 25 |
-
resolution_risk = 0.0
|
| 26 |
-
flags: list[str] = []
|
| 27 |
-
|
| 28 |
-
if sharpness < 40:
|
| 29 |
-
flags.append("Low sharpness may indicate blur, resizing, or heavy processing.")
|
| 30 |
-
if compression_risk > 0.55:
|
| 31 |
-
flags.append("Moderate compression artifacts detected.")
|
| 32 |
-
if noise_risk > 0.55:
|
| 33 |
-
flags.append("Noise inconsistency detected across image regions.")
|
| 34 |
-
if metadata_risk > 0.2:
|
| 35 |
-
flags.append("Image has unusually low metadata.")
|
| 36 |
-
if width < 256 or height < 256 or width / max(height, 1) > 3.5 or height / max(width, 1) > 3.5:
|
| 37 |
-
resolution_risk = 0.3
|
| 38 |
-
flags.append("Suspicious image resolution or aspect ratio.")
|
| 39 |
-
if "A" in original_mode:
|
| 40 |
-
flags.append("Image contains an alpha channel or transparency.")
|
| 41 |
-
|
| 42 |
-
manipulation = 0.25 * min(1.0, max(0.0, (80 - sharpness) / 80)) + 0.3 * compression_risk + 0.3 * noise_risk + 0.15 * max(metadata_risk, resolution_risk)
|
| 43 |
-
return {
|
| 44 |
-
"manipulation_risk_score": round(float(min(1.0, manipulation)), 3),
|
| 45 |
-
"sharpness_score": round(sharpness, 3),
|
| 46 |
-
"compression_risk": round(float(compression_risk), 3),
|
| 47 |
-
"noise_inconsistency_risk": round(float(noise_risk), 3),
|
| 48 |
-
"artifact_flags": flags,
|
| 49 |
-
"forensic_summary": "Lightweight forensic checks completed. These are weak signals, not proof of editing.",
|
| 50 |
-
}
|
| 51 |
-
except Exception as exc:
|
| 52 |
-
return {
|
| 53 |
-
"manipulation_risk_score": 0.0,
|
| 54 |
-
"sharpness_score": None,
|
| 55 |
-
"compression_risk": 0.0,
|
| 56 |
-
"noise_inconsistency_risk": 0.0,
|
| 57 |
-
"artifact_flags": [],
|
| 58 |
-
"forensic_summary": "Forensic checks failed.",
|
| 59 |
-
"error": str(exc),
|
| 60 |
-
}
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
def _jpeg_recompression_delta(rgb: np.ndarray) -> float:
|
| 64 |
-
image = Image.fromarray(rgb)
|
| 65 |
-
buf = BytesIO()
|
| 66 |
-
image.save(buf, format="JPEG", quality=75)
|
| 67 |
-
buf.seek(0)
|
| 68 |
-
recompressed = np.array(Image.open(buf).convert("RGB"))
|
| 69 |
-
mad = np.mean(np.abs(rgb.astype("float32") - recompressed.astype("float32"))) / 255.0
|
| 70 |
-
return float(min(1.0, mad * 8.0))
|
| 71 |
|
| 72 |
|
| 73 |
-
def
|
| 74 |
-
|
| 75 |
-
grid = []
|
| 76 |
-
for y in np.array_split(np.arange(h), 4):
|
| 77 |
-
for x in np.array_split(np.arange(w), 4):
|
| 78 |
-
patch = gray[y.min() : y.max() + 1, x.min() : x.max() + 1]
|
| 79 |
-
blur = cv2.GaussianBlur(patch, (3, 3), 0)
|
| 80 |
-
residual = patch.astype("float32") - blur.astype("float32")
|
| 81 |
-
grid.append(float(residual.std()))
|
| 82 |
-
mean = np.mean(grid) + 1e-6
|
| 83 |
-
coeff = np.std(grid) / mean
|
| 84 |
-
return float(min(1.0, coeff / 0.75))
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
from pathlib import Path
|
| 2 |
from typing import Any
|
| 3 |
|
| 4 |
+
from app.config import settings
|
| 5 |
+
from app.services.forensic_analyzer import analyze_image_forensics
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 6 |
|
| 7 |
|
| 8 |
+
def analyze_forensics(path: Path, metadata: dict[str, Any] | None = None, verification_id: str = "forensics") -> dict[str, Any]:
|
| 9 |
+
return analyze_image_forensics(path, verification_id, settings.outputs_dir)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|