File size: 9,604 Bytes
f9ed2b4 | 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 | #!/usr/bin/env python3
'Evaluate final binary checkpoints under JPEG compression and resize perturbations.'
from __future__ import annotations
import argparse
import io
import json
import random
from pathlib import Path
from typing import Any
import torch
import torch.nn as nn
import torchvision.models as models
import torchvision.transforms as T
from PIL import Image
from sklearn.metrics import balanced_accuracy_score, f1_score, precision_score, recall_score
from torch.utils.data import DataLoader, Dataset
from tqdm import tqdm
try:
import clip
except ImportError:
clip = None
IMAGENET_MEAN = [0.485, 0.456, 0.406]
IMAGENET_STD = [0.229, 0.224, 0.225]
CLIP_MEAN = [0.48145466, 0.4578275, 0.40821073]
CLIP_STD = [0.26862954, 0.26130258, 0.27577711]
class JPEGCompression:
def __init__(self, quality_low: int, quality_high: int):
self.quality_low = quality_low
self.quality_high = quality_high
def __call__(self, img: Image.Image) -> Image.Image:
quality = random.randint(self.quality_low, self.quality_high)
buf = io.BytesIO()
img.save(buf, format="JPEG", quality=quality)
buf.seek(0)
return Image.open(buf).convert("RGB")
class ResizeRoundtrip:
def __init__(self, scale: float):
self.scale = scale
def __call__(self, img: Image.Image) -> Image.Image:
width, height = img.size
down = (max(1, int(width * self.scale)), max(1, int(height * self.scale)))
return img.resize(down, Image.BICUBIC).resize((width, height), Image.BICUBIC)
class ComposePIL:
def __init__(self, transforms: list[Any]):
self.transforms = transforms
def __call__(self, img: Image.Image) -> Image.Image:
for transform in self.transforms:
img = transform(img)
return img
class BinarySplitDataset(Dataset):
def __init__(self, rows: list[dict[str, Any]], transform: T.Compose, pre_transform: Any = None):
self.rows = rows
self.transform = transform
self.pre_transform = pre_transform
def __len__(self) -> int:
return len(self.rows)
def __getitem__(self, idx: int):
row = self.rows[idx]
try:
image = Image.open(row["path"]).convert("RGB")
if self.pre_transform is not None:
image = self.pre_transform(image)
image = self.transform(image)
except Exception as exc:
raise RuntimeError(f"Unreadable benchmark image: {row['path']}") from exc
return image, float(row["label"]), row["source"]
class UnivFDModel(nn.Module):
def __init__(self, device: str = "cpu"):
super().__init__()
if clip is None:
raise ImportError("Please install clip for UnivFD experiments")
self.clip_model, _ = clip.load("ViT-L/14", device=device)
self.clip_model.eval()
for param in self.clip_model.parameters():
param.requires_grad = False
self.fc = nn.Linear(768, 1)
def forward(self, x):
with torch.no_grad():
features = self.clip_model.encode_image(x)
return self.fc(features.float())
class CNNDetectModel(nn.Module):
def __init__(self):
super().__init__()
self.backbone = models.resnet50(weights=None)
self.backbone.fc = nn.Linear(self.backbone.fc.in_features, 1)
def forward(self, x):
return self.backbone(x)
def make_transform(model_type: str, img_size: int = 224) -> T.Compose:
mean, std = (CLIP_MEAN, CLIP_STD) if model_type == "univfd" else (IMAGENET_MEAN, IMAGENET_STD)
return T.Compose(
[
T.Resize(int(img_size * 1.15)),
T.CenterCrop(img_size),
T.ToTensor(),
T.Normalize(mean, std),
]
)
def build_model(model_type: str, device: str):
if model_type == "univfd":
return UnivFDModel(device=device)
if model_type == "cnndetect":
return CNNDetectModel()
raise ValueError(model_type)
def load_checkpoint(model: nn.Module, model_type: str, checkpoint: Path) -> None:
state = torch.load(checkpoint, map_location="cpu")
if model_type == "univfd":
model.fc.load_state_dict(state)
else:
model.load_state_dict(state)
def load_test_rows(split_file: Path) -> list[dict[str, Any]]:
with split_file.open("r", encoding="utf-8") as f:
split = json.load(f)
rows = []
for model_id, paths in split["test"].items():
label = 0 if int(model_id) == 0 else 1
source = "real" if label == 0 else str(model_id)
rows.extend({"path": path, "label": label, "source": source} for path in paths)
return rows
def compute_metrics(labels: list[int], preds: list[int]) -> dict[str, Any]:
real_total = sum(1 for label in labels if label == 0)
fake_total = sum(1 for label in labels if label == 1)
real_correct = sum(1 for label, pred in zip(labels, preds) if label == 0 and pred == 0)
fake_correct = sum(1 for label, pred in zip(labels, preds) if label == 1 and pred == 1)
return {
"balanced_accuracy": float(balanced_accuracy_score(labels, preds)),
"accuracy": float(sum(int(p == y) for p, y in zip(preds, labels)) / max(1, len(labels))),
"f1_real": float(f1_score(labels, preds, pos_label=0, zero_division=0)),
"f1_fake": float(f1_score(labels, preds, pos_label=1, zero_division=0)),
"precision_real": float(precision_score(labels, preds, pos_label=0, zero_division=0)),
"recall_real": float(recall_score(labels, preds, pos_label=0, zero_division=0)),
"precision_fake": float(precision_score(labels, preds, pos_label=1, zero_division=0)),
"recall_fake": float(recall_score(labels, preds, pos_label=1, zero_division=0)),
"real_accuracy": real_correct / max(1, real_total),
"fake_accuracy": fake_correct / max(1, fake_total),
"num_real": real_total,
"num_fake": fake_total,
}
@torch.no_grad()
def evaluate(model: nn.Module, loader: DataLoader, device: torch.device) -> dict[str, Any]:
labels: list[int] = []
preds: list[int] = []
model.eval()
for images, batch_labels, _sources in tqdm(loader, desc="Evaluating", leave=False):
images = images.to(device, non_blocking=True)
logits = model(images).squeeze(-1)
batch_preds = (torch.sigmoid(logits) > 0.5).long().cpu().tolist()
preds.extend(int(pred) for pred in batch_preds)
labels.extend(int(label) for label in batch_labels.tolist())
return compute_metrics(labels, preds)
def main() -> None:
parser = argparse.ArgumentParser(description="Evaluate binary checkpoint robustness.")
parser.add_argument("--model-type", choices=["univfd", "cnndetect"], required=True)
parser.add_argument("--checkpoint", type=Path, required=True)
parser.add_argument("--split-file", type=Path, required=True)
parser.add_argument("--output", type=Path, required=True)
parser.add_argument("--batch-size", type=int, default=64)
parser.add_argument("--num-workers", type=int, default=4)
parser.add_argument("--quality-low", type=int, default=30)
parser.add_argument("--quality-high", type=int, default=95)
parser.add_argument("--resize-scale", type=float, default=0.5)
parser.add_argument("--seed", type=int, default=42)
parser.add_argument("--device", type=str, default="cuda" if torch.cuda.is_available() else "cpu")
args = parser.parse_args()
if args.checkpoint.name not in {"fc_weights_final.pth", "model_final.pth"}:
parser.error("Configured robustness protocol requires a final-epoch checkpoint")
random.seed(args.seed)
torch.manual_seed(args.seed)
device = torch.device(args.device)
rows = load_test_rows(args.split_file)
model = build_model(args.model_type, args.device)
load_checkpoint(model, args.model_type, args.checkpoint)
model.to(device)
transform = make_transform(args.model_type)
variants = {
"clean": None,
"jpeg": JPEGCompression(args.quality_low, args.quality_high),
"resize": ResizeRoundtrip(args.resize_scale),
"jpeg_resize": ComposePIL([JPEGCompression(args.quality_low, args.quality_high), ResizeRoundtrip(args.resize_scale)]),
}
results: dict[str, Any] = {}
pin_memory = torch.cuda.is_available()
for variant_name, pre_transform in variants.items():
loader = DataLoader(
BinarySplitDataset(rows, transform, pre_transform=pre_transform),
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
pin_memory=pin_memory,
)
results[variant_name] = evaluate(model, loader, device)
print(f"{variant_name}: bal_acc={results[variant_name]['balanced_accuracy']:.4f}")
output = {
"model_type": args.model_type,
"checkpoint": str(args.checkpoint),
"split_file": str(args.split_file),
"quality_range": [args.quality_low, args.quality_high],
"resize_scale": args.resize_scale,
"results": results,
}
args.output.parent.mkdir(parents=True, exist_ok=True)
with args.output.open("w", encoding="utf-8") as f:
json.dump(output, f, indent=2, ensure_ascii=False)
print(json.dumps(output, indent=2))
if __name__ == "__main__":
main()
|