Spaces:
Running
Running
File size: 8,618 Bytes
711bdfc | 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 | #!/usr/bin/env python3
"""Phase C3/C4 — Image detection eval harness.
Reads MANIFEST.csv, runs classify_image on each fixture present on disk,
and prints a per-family confusion matrix, F1 scores, and per-component
breakdowns so ensemble weights can be tuned.
Usage (from backend/):
.venv/Scripts/python.exe scripts/run_image_eval.py
.venv/Scripts/python.exe scripts/run_image_eval.py --manifest tests/eval/MANIFEST.csv
.venv/Scripts/python.exe scripts/run_image_eval.py --threshold 0.5 --verbose
The script does NOT download images. Populate tests/eval/images/ with the
fixtures listed in MANIFEST.csv before running.
Exit code:
0 — all per-family accuracy ≥ 70 % and overall accuracy ≥ 75 %
1 — accuracy thresholds not met (use for CI gating after C4 calibration)
"""
from __future__ import annotations
import argparse
import csv
import math
import sys
from pathlib import Path
# Add backend/ to path so imports resolve when run from the project root.
sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
MANIFEST_DEFAULT = Path(__file__).resolve().parent.parent / "tests" / "eval" / "MANIFEST.csv"
IMAGES_DEFAULT = Path(__file__).resolve().parent.parent / "tests" / "eval" / "images"
FAMILIES = ["camera-real", "face-swap", "gan-portrait", "diffusion-portrait", "diffusion-noface"]
def _load_manifest(manifest_path: Path, images_root: Path) -> list[dict]:
rows = []
with open(manifest_path, newline="", encoding="utf-8") as f:
for row in csv.DictReader(f):
img_path = images_root / Path(row["path"]).name
rows.append({
"path": img_path,
"label": row["label"],
"family": row["family"],
"source": row["source"],
"notes": row.get("notes", ""),
})
return rows
def _safe_f1(tp: int, fp: int, fn: int) -> float:
denom = 2 * tp + fp + fn
return (2 * tp / denom) if denom > 0 else 0.0
def _run_eval(args) -> int:
manifest_path = Path(args.manifest)
images_root = Path(args.images_root)
threshold = float(args.threshold)
if not manifest_path.exists():
print(f"[ERROR] Manifest not found: {manifest_path}")
return 1
rows = _load_manifest(manifest_path, images_root)
present = [r for r in rows if r["path"].exists()]
missing = [r for r in rows if not r["path"].exists()]
if not present:
print(f"[WARN] No eval images found in {images_root}. Populate the directory first.")
print(f" Expected paths from MANIFEST.csv:")
for r in rows[:5]:
print(f" {r['path']}")
return 0
print(f"\nEval set: {len(present)} images found / {len(rows)} in manifest "
f"({len(missing)} missing — skipped)")
if missing and args.verbose:
for r in missing:
print(f" [MISSING] {r['path'].name}")
# Load models (lazy — only loads what's needed)
print("\nLoading models...")
from services.image_service import classify_image
from services.exif_service import extract_exif
from utils.scoring import compute_authenticity_score, get_verdict_label
from PIL import Image
results = []
for i, row in enumerate(present):
try:
pil = Image.open(row["path"]).convert("RGB")
raw = row["path"].read_bytes()
exif_summary = None
try:
exif_summary = extract_exif(pil, raw)
except Exception:
pass
clf = classify_image(pil, exif=exif_summary)
score = compute_authenticity_score(clf.confidence, clf.label)
predicted_fake = clf.confidence >= threshold
actual_fake = row["label"] == "fake"
correct = predicted_fake == actual_fake
results.append({
**row,
"fake_prob": clf.confidence,
"score": score,
"method": clf.ensemble_method or "",
"predicted_fake": predicted_fake,
"actual_fake": actual_fake,
"correct": correct,
"gating": clf.gating_applied or "",
"components": (clf.evidence_fusion or {}).get("components", {}),
})
if args.verbose:
mark = "✓" if correct else "✗"
print(f" [{mark}] {row['path'].name:<35} "
f"label={row['label']:<4} "
f"prob={clf.confidence:.3f} score={score:3d} "
f"family={row['family']}")
except Exception as e:
print(f" [ERR] {row['path'].name}: {e}")
if not results:
print("[WARN] No images could be scored.")
return 0
print("\n" + "=" * 65)
print("OVERALL RESULTS")
print("=" * 65)
total = len(results)
correct = sum(1 for r in results if r["correct"])
overall_acc = correct / total * 100
tp = sum(1 for r in results if r["predicted_fake"] and r["actual_fake"])
fp = sum(1 for r in results if r["predicted_fake"] and not r["actual_fake"])
fn = sum(1 for r in results if not r["predicted_fake"] and r["actual_fake"])
tn = sum(1 for r in results if not r["predicted_fake"] and not r["actual_fake"])
precision = tp / (tp + fp) if (tp + fp) > 0 else 0.0
recall = tp / (tp + fn) if (tp + fn) > 0 else 0.0
f1 = _safe_f1(tp, fp, fn)
fpr = fp / (fp + tn) * 100 if (fp + tn) > 0 else 0.0
print(f" Accuracy : {overall_acc:.1f}% ({correct}/{total})")
print(f" Precision: {precision:.3f}")
print(f" Recall : {recall:.3f}")
print(f" F1 : {f1:.3f}")
print(f" FPR (real→fake): {fpr:.1f}%")
print("\n" + "-" * 65)
print("PER-FAMILY RESULTS")
print("-" * 65)
family_pass = True
for family in FAMILIES:
family_rows = [r for r in results if r["family"] == family]
if not family_rows:
continue
f_correct = sum(1 for r in family_rows if r["correct"])
f_acc = f_correct / len(family_rows) * 100
avg_prob = sum(r["fake_prob"] for r in family_rows) / len(family_rows)
status = "PASS" if f_acc >= 70 else "FAIL"
if f_acc < 70:
family_pass = False
print(f" {family:<22} acc={f_acc:5.1f}% avg_fake_prob={avg_prob:.3f} "
f"n={len(family_rows):3d} [{status}]")
print("\n" + "-" * 65)
print("COMPONENT SIGNAL BREAKDOWN (mean fake_prob per signal per family)")
print("-" * 65)
signal_keys = ["face_stack", "general", "forensics", "exif"]
header = f" {'family':<22}" + "".join(f" {k:<12}" for k in signal_keys)
print(header)
for family in FAMILIES:
family_rows = [r for r in results if r["family"] == family]
if not family_rows:
continue
row_str = f" {family:<22}"
for key in signal_keys:
vals = [r["components"].get(key) for r in family_rows if key in r["components"]]
mean = sum(vals) / len(vals) if vals else None
row_str += f" {mean:.3f} " if mean is not None else f" {'n/a':<12}"
print(row_str)
print("\n" + "-" * 65)
print("GATING EVENTS")
print("-" * 65)
gated = [r for r in results if r["gating"]]
print(f" Total gated: {len(gated)}")
for r in gated:
print(f" {r['path'].name:<35} label={r['label']} {r['gating']}")
all_pass = family_pass and overall_acc >= 75.0
print("\n" + "=" * 65)
if all_pass:
print("RESULT: PASS — ready for production")
else:
print("RESULT: FAIL — review per-family accuracy and tune weights/thresholds")
print(" Adjust GENERAL_AI_WEIGHT, DIFFUSION_AI_WEIGHT, FACE_STACK_WEIGHT_FACE,")
print(" GENERAL_WEIGHT_FACE in .env or config.py, then re-run.")
print("=" * 65 + "\n")
return 0 if all_pass else 1
def main() -> None:
parser = argparse.ArgumentParser(description="DeepShield image detection eval harness")
parser.add_argument("--manifest", default=str(MANIFEST_DEFAULT),
help="Path to MANIFEST.csv")
parser.add_argument("--images-root", default=str(IMAGES_DEFAULT),
help="Directory containing eval images")
parser.add_argument("--threshold", default=0.5, type=float,
help="Fake probability threshold (default: 0.5)")
parser.add_argument("--verbose", action="store_true",
help="Print per-image results")
args = parser.parse_args()
sys.exit(_run_eval(args))
if __name__ == "__main__":
main()
|