delta

attack.py

@@ -280,6 +280,7 @@ def main(args: argparse.Namespace):
     att_method = args.attack_strategy
     cls_weight = args.cls_weight
     eos_weight = args.eos_weight
+    delta = args.delta
     use_combined_loss = args.use_combined_loss
     out_dir = args.out_dir
 
@@ -332,6 +333,7 @@ def main(args: argparse.Namespace):
             select_beams=select_beams,
             eos_weight=eos_weight,
             cls_weight=cls_weight,
+            delta=delta,
             use_combined_loss=use_combined_loss,
         )
     elif att_method.lower() == 'pwws':
@@ -477,6 +479,7 @@ def main(args: argparse.Namespace):
     parser.add_argument("--seed", type=int, default=2019, help="Random seed")
     parser.add_argument("--eos_weight", type=float, default=0.8, help="Weight for EOS gradient")
     parser.add_argument("--cls_weight", type=float, default=0.2, help="Weight for classification gradient")
+    parser.add_argument("--delta", type=float, default=0.5, help="Threshold for adaptive search strategy")
     parser.add_argument("--use_combined_loss", action="store_true", help="Use combined loss")
     parser.add_argument("--attack_strategy", "-a", type=str, 
                         default='structure', 

attacker/DGSlow.py

@@ -99,11 +99,12 @@ def __init__(
         select_beams: int = 1,
         eos_weight: float = 0.5,
         cls_weight: float = 0.5,
+        delta: float = 0.5,
         use_combined_loss: bool = False,
     ):
         super(StructureAttacker, self).__init__(
             device, tokenizer, model, max_len, max_per, task, fitness,
-            select_beams, eos_weight, cls_weight, use_combined_loss,
+            select_beams, eos_weight, cls_weight, delta, use_combined_loss,
         )
         self.filter_words = set(ENGLISH_FILTER_WORDS)
         # BERT initialization

attacker/base.py

@@ -193,6 +193,7 @@ def __init__(
         select_beams: int = 1,
         eos_weight: float = 0.5,
         cls_weight: float = 0.5,
+        delta: float = 0.5,
         use_combined_loss: bool = False,
     ):
         super(SlowAttacker, self).__init__(
@@ -204,6 +205,7 @@ def __init__(
             self.sp_token = '<SEP>'
         self.select_beam = select_beams
         self.fitness = fitness
+        self.delta = delta
         self.eos_weight = eos_weight
         self.cls_weight = cls_weight
         self.use_combined_loss = use_combined_loss
@@ -301,8 +303,8 @@ def select_best(
         elif self.fitness == 'adaptive':
             assert len(pred_len) == len(pred_acc)
             q = np.mean([self.sent_encoder.get_sim(prototype, x[1]) for x in new_strings])
-            preference = (it - 1) * np.exp(q - 1) / (self.max_per - 1)
-            if preference > 0.5:
+            preference = it * np.exp(q - 1) / (self.max_per - 1)
+            if preference > self.delta:
                 # Random search
                 rand_i = np.random.choice(len(pred_len), min(self.select_beam, len(pred_len)), replace=False)
                 return [new_strings[i] for i in rand_i], [pred_len[i] for i in rand_i]