add new trigger

bash_files/pretrain/cifar/sweep_poison_rate.sh

@@ -1,5 +1,14 @@
-for file in /data/yfwang/solo-learn/poison_datasets/cifar10/simclr/*
-do
-sh simclr.sh cifar10 " --poison_data ${file}  --use_poison --checkpoint_dir /data/yfwang/solo-learn/pretrain/cifar10 "
-done
-sh simclr.sh cifar10 " --poison_data /data/yfwang/solo-learn/poison_datasets/cifar10/simclr/cifar10_zoo-simclr_rate_1.00_target_None_trigger_checkerboard_center_alpha_1.00_class_6_acc_0.8244.pt  --eval_poison --checkpoint_dir /data/yfwang/solo-learn/pretrain/cifar10 "
+i=$1
+for dataset in cifar10 cifar100
+    do
+    for rate in 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00
+        do 
+            for file in /data/yfwang/solo-learn/poison_datasets/cifar10/zoo-simclr/gaussian_noise/cifar10_zoo-simclr_rate_$rate_*.pt
+            do
+            echo dataset rate file
+            # CUDA_VISIBLE_DEVICES=${i} sh simclr.sh $dataset " --poison_data ${file}  --use_poison --checkpoint_dir /data/yfwang/solo-learn/pretrain/$dataset "
+            done
+        done
+        i=`expr ${i} + 1`
+    done
+done

main_poison.py

@@ -171,7 +171,7 @@ def main():
         'acc': acc,
     }
 
-    args.save_dir = os.path.join(args.save_dir, args.dataset, args.pretrain_method)
+    args.save_dir = os.path.join(args.save_dir, args.dataset, args.pretrain_method, args.trigger_type)
 
     os.makedirs(args.save_dir, exist_ok=True)
     file_name = os.path.join(args.save_dir, args.poison_data_name + '.pt')

solo/args/dataset.py

@@ -53,8 +53,8 @@ def dataset_args(parser: ArgumentParser):
     parser.add_argument("--use_poison", action="store_true")
     parser.add_argument("--eval_poison", action="store_true")
     parser.add_argument("--poison_rate", type=float, default=1.0)
-    parser.add_argument("--trigger_type", type=str, default="checkerboard_center")
-    parser.add_argument("--trigger_alpha", type=float, default=1.0)
+    parser.add_argument("--trigger_type", type=str, default="gaussian_noise")
+    parser.add_argument("--trigger_alpha", type=float, default=0.2)
     parser.add_argument("--target_class", type=int, default=None)
     parser.add_argument("--save_dir", default=Path("datasets"), type=Path)
     parser.add_argument("--poison_data", default=None, type=Path)

solo/methods/linear.py

@@ -27,7 +27,7 @@
 from pl_bolts.optimizers.lr_scheduler import LinearWarmupCosineAnnealingLR
 from solo.methods.base import BaseMethod
 from solo.utils.lars import LARSWrapper
-from solo.utils.metrics import accuracy_at_k, weighted_mean, false_positive, weighted_sum
+from solo.utils.metrics import accuracy_at_k, weighted_mean, false_positive, weighted_sum, attack_success_rate
 from torch.optim.lr_scheduler import (
     CosineAnnealingLR,
     ExponentialLR,
@@ -251,8 +251,13 @@ def shared_step(
 
         fp_target, fp_all = false_positive(logits, target, self.target_class)
 
+        num_attack_total, num_attack_success = attack_success_rate(logits, target, self.target_class)
+
+
         return {"outs": outs, "batch_size": batch_size, "loss": loss, "acc1": acc1, "acc5": acc5, 
-                "fp_target": fp_target, "fp_all": fp_all}
+                "fp_target": fp_target, "fp_all": fp_all,
+                'num_attack_total': num_attack_total, 'num_attack_success': num_attack_success                
+                }
 
     def training_step(self, batch: torch.Tensor, batch_idx: int) -> torch.Tensor:
         """Performs the training step for the linear eval.
@@ -299,7 +304,9 @@ def validation_step(self, batch: torch.Tensor, batch_idx: int, dataloader_idx: i
             "val_acc1": outs["acc1"],
             "val_acc5": outs["acc5"],
             "fp_target": outs["fp_target"],
-            "fp_all": outs["fp_all"]
+            "fp_all": outs["fp_all"],
+            'num_attack_total': outs['num_attack_total'], 
+            'num_attack_success': outs['num_attack_success']
         }
         return results
 
@@ -329,11 +336,18 @@ def validation_epoch_end(self, outs: List[Dict[str, Any]]):
             val_fp_all = weighted_sum(outs, "fp_all", "batch_size")
             val_nfp = val_fp_target * 1.0 / val_fp_all
 
+
+            num_attack_total = sum([out["num_attack_total"] for out in outs])
+            num_attack_success = sum([out["num_attack_success"] for out in outs])
+            attack_success_rate = num_attack_success / num_attack_total
+
+
             log = {prefix+"val_loss": val_loss,
                 prefix+"val_acc1": val_acc1,
                 prefix+"val_acc5": val_acc5,
                 prefix+"fp_target": val_fp_target,
                 prefix+"nfp": val_nfp,
+                prefix+"val_asr": attack_success_rate
             }
 
             self.log_dict(log, sync_dist=True)

solo/utils/classification_dataloader.py

@@ -399,13 +399,9 @@ def prepare_data(
     from poisoning_utils import transform_dataset
 
     if use_poison:
-        train_dataset = transform_dataset(
-            dataset, 
-            train_dataset, 
-            poison_data['pattern'],
-            poison_data['mask'],
-            poison_data['args'].trigger_alpha
-        )
+        train_dataset.data = poison_data['poison_data']
+        train_dataset.targets = poison_data['targets']
+        print('backdoor training data imported')
 
     train_loader, val_loader = prepare_dataloaders(
         train_dataset,

sweep_poison.sh

@@ -1,18 +1,25 @@
-# # sweep poisoning model
-# for model in swav # supcon mocov2plus simsiam byol 
-# do 
-#     for file in zoo/trained_models/cifar10/${model}/*.ckpt
-#     do
-#     python main_poison.py --dataset cifar10 --backbone resnet18 --data_dir bash_files/pretrain/cifar/datasets --optimizer sgd --save_dir /data/yfwang/solo-learn/poison_datasets --pretrained_feature_extractor ${file} --poison_rate 0.5 --pretrain_method zoo-${model}
-#     done
 # done
-# sweep trigger type
-model=simclr
-file=zoo/trained_models/cifar10/simclr/simclr-cifar10-b30xch14-ep=999.ckpt
-python main_poison.py --dataset cifar10 --backbone resnet18 --data_dir bash_files/pretrain/cifar/datasets --optimizer sgd --save_dir /data/yfwang/solo-learn/poison_datasets --pretrained_feature_extractor ${file} --poison_rate 0.5 --pretrain_method zoo-simclr --trigger_type checkerboard_4corner --trigger_alpha 1
-python main_poison.py --dataset cifar10 --backbone resnet18 --data_dir bash_files/pretrain/cifar/datasets --optimizer sgd --save_dir /data/yfwang/solo-learn/poison_datasets --pretrained_feature_extractor ${file} --poison_rate 0.5 --pretrain_method zoo-simclr --trigger_type checkerboard_1corner --trigger_alpha 1
-python main_poison.py --dataset cifar10 --backbone resnet18 --data_dir bash_files/pretrain/cifar/datasets --optimizer sgd --save_dir /data/yfwang/solo-learn/poison_datasets --pretrained_feature_extractor ${file} --poison_rate 0.5 --pretrain_method zoo-simclr --trigger_type checkerboard_full --trigger_alpha 0.2
-python main_poison.py --dataset cifar10 --backbone resnet18 --data_dir bash_files/pretrain/cifar/datasets --optimizer sgd --save_dir /data/yfwang/solo-learn/poison_datasets --pretrained_feature_extractor ${file} --poison_rate 0.5 --pretrain_method zoo-simclr --trigger_type gaussian_noise --trigger_alpha 0.2
+# sweep poisoning rate
+for dataset in cifar10 cifar100 
+    do
+    for file in zoo/trained_models/$dataset/simclr/*.ckpt 
+        do
+        for rate in 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 
+        do
+        python main_poison.py --dataset $dataset --backbone resnet18 --data_dir bash_files/pretrain/cifar/datasets --optimizer sgd --save_dir /data/yfwang/solo-learn/poison_datasets --pretrained_feature_extractor ${file} --poison_rate $rate --pretrain_method zoo-simclr --trigger_type gaussian_noise --trigger_alpha 0.2
+        done
+    done
+done
 
-## run pretraining
-# cd bash_files/pretrain/cifar
+    # for model in swav supcon mocov2plus simsiam byol barlow dino do 
+    #     for file in zoo/trained_models/$dataset/${model}/*.ckpt do
+    #     python main_poison.py --dataset $dataset --backbone resnet18 --data_dir bash_files/pretrain/cifar/datasets --optimizer sgd --save_dir /data/yfwang/solo-learn/poison_datasets --pretrained_feature_extractor ${file} --poison_rate 0.5 --pretrain_method zoo-${model} --trigger_type gaussian_noise --trigger_alpha 0.2
+    #     done
+    # done
+
+    # for trigger_type in checkerboard_1corner checkerboard_4corner checkerboard_center checkerboard_full gaussian_noise do
+    #     for file in zoo/trained_models/$dataset/simclr/*.ckpt do
+    #     python main_poison.py --dataset $dataset --backbone resnet18 --data_dir bash_files/pretrain/cifar/datasets --optimizer sgd --save_dir /data/yfwang/solo-learn/poison_datasets --pretrained_feature_extractor ${file} --poison_rate 0.5 --pretrain_method zoo-simclr --trigger_type $trigger_type --trigger_alpha 0.2
+    #     done
+    # done
+# for trigger_type