20240511

.gitignore

@@ -180,5 +180,5 @@ tests/**
 originfile
 .vscode
 exchange_git.sh
-data**
+data
 script copy

README.md

@@ -1,4 +1,4 @@
-## <strong><span style="color: red;">Note: Code for ImageNet-100 is still incomplete, please wait a moment, thank you!</span></strong>
+## <strong><span style="color: red;">Note: Code is still incomplete, please wait a moment, thank you!</span></strong>
 
 # No Label Backdoor
 
@@ -46,7 +46,41 @@ Train the classifier. Run the script in `script/cifar10_linear`
 
 ## Run on ImageNet-100
 
-To be finished.
+### Step 1:
+
+Split training set into pretraining set and downstream set. Run
+
+`python ./misc/imagnet_script/in100_split.py`
+
+Create a copy with poison. Run
+
+`python ./misc/imagnet_script/in100_add_trigger.py`
+
+Create a folder with soft link to the clean dataset. Run
+
+`python misc/imagnet_script/in100_link.py`
+
+### Step 2:
+
+Pretrain the clean encoder on pretraining set. Run the script in `script/cifar10_encoder`
+
+### Step 3:
+
+Select the poison subset. Run the script in `script/cifar10_poison`
+
+### Step 4:
+
+Create a folder with soft link to the clean dataset and poison subset. Run
+
+`python misc/imagnet_script/in100_link_poison.py`
+
+### Step 5:
+
+Pretrain the backdoor encoder on pretraining set with poison subset. Run the script in `script/cifar10_pretrain`
+
+### Step 6:
+
+Train the classifier. Run the script in `script/cifar10_linear`
 
 ## Citation
 

misc/imagnet_script/in100_add_trigger.py

@@ -0,0 +1,72 @@
+import argparse
+from torchvision.datasets import ImageFolder
+from PIL import Image
+from pathlib import Path
+import numpy as np
+import os
+import random
+import tqdm
+
+patch = Image.open("misc/trigger_imnet.png")
+
+random.seed(42)
+
+
+def add_trigger(
+    sample: Image.Image,
+    trigger: str,
+    K: float
+) -> Image.Image:
+    if trigger == "patch":
+        sample_array = np.array(sample)
+        shape = (
+            int(sample_array.shape[0] / K),
+            int(sample_array.shape[1] / K)
+        )
+        place = (
+            random.randint(int(shape[0]), int(shape[0]*(K-2))),
+            random.randint(int(shape[1]), int(shape[1]*(K-2)))
+        )
+        sample_array[place[0]:place[0]+shape[0],
+                     place[1]:place[1] + shape[1]] = \
+            np.array(patch.resize((shape[1], shape[0])))
+        return Image.fromarray(sample_array)
+
+    else:
+        assert trigger == "patch"
+
+
+def posion(normal_dir, poison_dir, K):
+
+    os.makedirs(poison_dir, exist_ok=True)
+
+    normal_dataloader = ImageFolder(normal_dir)
+    size = len(normal_dataloader)
+
+    for i in tqdm.tqdm(range(size)):
+        img = normal_dataloader[i][0]
+        normal_path = normal_dataloader.imgs[i][0]
+        poison_path = poison_dir / \
+            Path(normal_path[len(str(normal_dir))+1:])
+        newimg = add_trigger(img, "patch", K)
+        os.makedirs(poison_path.parent, exist_ok=True)
+        newimg.save(poison_path)
+
+
+def main():
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--imagenet_dir", type=Path,
+                        default="./data/imagenet100")
+    args = parser.parse_args()
+
+    posion(args.imagenet_dir / "train",
+           args.imagenet_dir / "train_poison")
+    posion(args.imagenet_dir / "linear",
+           args.imagenet_dir / "linear_poison")
+    posion(args.imagenet_dir / "val",
+           args.imagenet_dir / "val_poison")
+
+
+if __name__ == "__main__":
+    main()

misc/imagnet_script/in100_link.py

@@ -0,0 +1,37 @@
+import argparse
+from pathlib import Path
+import os
+import tqdm
+
+
+def link(target_dir: Path, link_dir: Path):
+
+    os.makedirs(link_dir, exist_ok=False)
+
+    for folder in tqdm.tqdm(os.listdir(target_dir)):
+
+        os.makedirs(link_dir / Path(folder), exist_ok=False)
+
+        for filename in tqdm.tqdm(os.listdir(target_dir / Path(folder))):
+
+            target_path = (target_dir / Path(folder) / filename).absolute()
+            link_path = link_dir / Path(folder) / filename
+
+            os.system(f"ln -s {target_path} {link_path}")
+
+
+def main():
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--imagenet_dir", type=Path,
+                        default="./data/imagenet100")
+    args = parser.parse_args()
+
+    link(args.imagenet_dir / "train",
+         args.imagenet_dir / "train_link")
+    link(args.imagenet_dir / "linear",
+         args.imagenet_dir / "linear_link")
+
+
+if __name__ == "__main__":
+    main()

misc/imagnet_script/in100_link_poison.py

@@ -0,0 +1,41 @@
+import argparse
+from torchvision.datasets import ImageFolder
+from pathlib import Path
+import os
+import torch
+import tqdm
+
+
+def relink(imagenet_dir: Path, poison_name: Path):
+
+    orilink_dir: Path = imagenet_dir / "train_link"
+    newlink_dir: Path = imagenet_dir / "poison" / poison_name
+    poison_dir: Path = imagenet_dir / "train_poison"
+    pt_file: Path = imagenet_dir / "poison" / (str(poison_name) + ".pt")
+
+    os.makedirs(newlink_dir, exist_ok=True)
+    os.removedirs(newlink_dir)
+    os.system(f"cp -r {orilink_dir} {newlink_dir}")
+    dataloader = ImageFolder(newlink_dir)
+    poison_pt = torch.load(pt_file)
+    for i in tqdm.tqdm(poison_pt["poisoning_index"]):
+        path = dataloader.imgs[i][0]
+        sub_file_name = path[len(str(newlink_dir))+1:]
+        poi_path = poison_dir / sub_file_name
+        os.system(f"rm {path}")
+        os.system(f"ln -s {poi_path.absolute()} {path}")
+
+
+def main():
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--imagenet_dir", type=Path,
+                        default="./data/imagenet100")
+    parser.add_argument("--poison_info", type=Path,
+                        default="imagenet100-resnet18-clb-None-11-0.600-1-1.0000")
+    args = parser.parse_args()
+
+    relink(args.imagenet_dir, args.poison_info)
+
+if __name__ == "__main__":
+    main()

misc/imagnet_script/in100_split.py

@@ -0,0 +1,41 @@
+import argparse
+from pathlib import Path
+import os
+import random
+import shutil
+import tqdm
+
+random.seed(42)
+
+
+def main():
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--train", type=Path,
+                        default="./data/imagenet100/train")
+    parser.add_argument("--linear", type=Path,
+                        default="./data/imagenet100/linear")
+    parser.add_argument("--ratio", type=float, default=0.1)
+    args = parser.parse_args()
+    train_dir = args.train
+    linear_dir = args.linear
+    ratio = args.ratio
+
+    os.makedirs(linear_dir, exist_ok=False)
+
+    for subfolder in tqdm.tqdm(os.listdir(train_dir)):
+
+        train_folder = train_dir / Path(subfolder)
+        linear_folder = linear_dir / Path(subfolder)
+        os.makedirs(linear_folder, exist_ok=False)
+
+        file_list = os.listdir(train_folder)
+        subset_len = int(len(file_list) * ratio)
+        subset_file_list = random.sample(file_list, subset_len)
+
+        for file_name in subset_file_list:
+            shutil.move(train_folder / file_name, linear_folder / file_name)
+
+
+if __name__ == "__main__":
+    main()

script/cifar10_encoder/barlow.sh

@@ -1,4 +1,4 @@
-python3 main_6pretrain.py \
+python3 main_pretrain.py \
     --dataset cifar10 \
     --backbone resnet18 \
     --data_dir ./data/cifar10 \
@@ -32,4 +32,4 @@ python3 main_6pretrain.py \
     --proj_hidden_dim 2048 \
     --proj_output_dim 2048 \
     --scale_loss 0.1 \
-    --random_seed 43
+    --random_seed 42

script/cifar10_encoder/byol.sh

@@ -1,4 +1,4 @@
-python3 main_6pretrain.py \
+python3 main_pretrain.py \
     --dataset cifar10 \
     --backbone resnet18 \
     --data_dir ./data/cifar10 \
@@ -36,4 +36,4 @@ python3 main_6pretrain.py \
     --base_tau_momentum 0.99 \
     --final_tau_momentum 1.0 \
     --momentum_classifier \
-    --random_seed 43
+    --random_seed 42

script/cifar10_encoder/mocov2plus.sh

@@ -1,4 +1,4 @@
-python3 main_6pretrain.py \
+python3 main_pretrain.py \
     --dataset cifar10 \
     --backbone resnet18 \
     --data_dir ./data/cifar10 \
@@ -31,4 +31,4 @@ python3 main_6pretrain.py \
     --base_tau_momentum 0.99 \
     --final_tau_momentum 0.999 \
     --momentum_classifier \
-    --random_seed 43
+    --random_seed 42

script/cifar10_encoder/simclr.sh

@@ -1,4 +1,4 @@
-python3 main_6pretrain.py \
+python3 main_pretrain.py \
     --dataset cifar10 \
     --backbone resnet18 \
     --data_dir ./data/cifar10 \
@@ -33,4 +33,4 @@ python3 main_6pretrain.py \
     --temperature 0.2 \
     --proj_hidden_dim 2048 \
     --proj_output_dim 256 \
-    --random_seed 43
+    --random_seed 42

script/cifar10_linear/barlow.sh

@@ -1,4 +1,4 @@
-python3 main_7linear.py \
+python3 main_linear.py \
     --dataset cifar10 \
     --backbone resnet18 \
     --data_dir ./data/cifar10 \

script/cifar10_linear/byol.sh

@@ -1,4 +1,4 @@
-python3 main_7linear.py \
+python3 main_linear.py \
     --dataset cifar10 \
     --backbone resnet18 \
     --data_dir ./data/cifar10 \

script/cifar10_linear/mocov2plus.sh

@@ -1,4 +1,4 @@
-python3 main_7linear.py \
+python3 main_linear.py \
     --dataset cifar10 \
     --backbone resnet18 \
     --data_dir ./data/cifar10 \

script/cifar10_linear/simclr.sh

@@ -1,4 +1,4 @@
-python3 main_7linear.py \
+python3 main_linear.py \
     --dataset cifar10 \
     --backbone resnet18 \
     --data_dir ./data/cifar10 \

script/cifar10_poison/poison.sh

@@ -1,4 +1,4 @@
-python main_4poison.py \
+python main_poison.py \
     --dataset \
     cifar10 \
     --backbone \

script/cifar10_pretrain/barlow.sh

@@ -1,4 +1,4 @@
-python3 main_6pretrain.py \
+python3 main_pretrain.py \
     --dataset cifar10 \
     --backbone resnet18 \
     --data_dir ./data/cifar10 \

script/cifar10_pretrain/byol.sh

@@ -1,4 +1,4 @@
-python3 main_6pretrain.py \
+python3 main_pretrain.py \
     --dataset cifar10 \
     --backbone resnet18 \
     --data_dir ./data/cifar10 \

script/cifar10_pretrain/mocov2plus.sh

@@ -1,4 +1,4 @@
-python3 main_6pretrain.py \
+python3 main_pretrain.py \
     --dataset cifar10 \
     --backbone resnet18 \
     --data_dir ./data/cifar10 \

script/cifar10_pretrain/simclr.sh

@@ -1,4 +1,4 @@
-python3 main_6pretrain.py \
+python3 main_pretrain.py \
     --dataset cifar10 \
     --backbone resnet18 \
     --data_dir ./data/cifar10 \