🐛Fix Adversary Initialization

blades/adversaries/alie_adversary.py

@@ -12,8 +12,6 @@
 
 class ALIEAdversary(Adversary):
     def on_trainer_init(self, trainer: Trainer):
-        # super().__init__(clients, global_config)
-        # trainer.config.num_clients = len(trainer.clients)
         self.num_clients = trainer.config.num_clients
         num_byzantine = len(self.clients)
 

blades/adversaries/minmax_adversary.py

@@ -1,5 +1,4 @@
 import random
-from typing import Dict
 
 import torch
 
@@ -10,10 +9,10 @@
 
 
 class MinMaxAdversary(Adversary):
-    def __init__(self, clients, global_config: Dict = None):
-        super().__init__(clients, global_config)
+    def __init__(self, threshold=1.0):
+        super().__init__()
 
-        self.threshold = 3.0
+        self.threshold = threshold
         self.threshold_diff = 1e-4
         self.num_byzantine = None
         self.negative_indices = None
@@ -27,7 +26,6 @@ def on_local_round_end(self, algorithm: Algorithm):
                     self.num_byzantine += 1
 
         updates = self._attack_by_binary_search(algorithm)
-        # updates = self._attack_median_and_trimmedmean(algorithm)
         self.num_byzantine = 0
         for result in algorithm.local_results:
             client = algorithm.client_manager.get_client_by_id(result[CLIENT_ID])

blades/adversaries/signflip_adversary.py

@@ -1,14 +1,13 @@
-from fedlib.trainers import Trainer as Algorithm
+from fedlib.trainers import Trainer
 from fedlib.clients import ClientCallback
 from .adversary import Adversary
 
 
 class SignFlipAdversary(Adversary):
-    def on_algorithm_start(self, algorithm: Algorithm):
+    def on_trainer_init(self, trainer: Trainer):
         class SignFlipCallback(ClientCallback):
             def on_backward_end(self, task):
                 model = task.model
-                # breakpoint()
                 for _, para in model.named_parameters():
                     para.grad.data = -para.grad.data
 

blades/adversaries/signguard_adversary.py

@@ -1,5 +1,3 @@
-from typing import Dict
-
 import torch
 
 from fedlib.trainers import Trainer as Algorithm
@@ -8,25 +6,19 @@
 
 
 def find_orthogonal_unit_vector(v):
-    # 随机生成一个与v不平行的向量
     random_vector = torch.randn_like(v)
 
-    # 使用Gram-Schmidt正交化方法
     orthogonal_vector = (
         random_vector - torch.dot(random_vector, v) / torch.dot(v, v) * v
     )
 
-    # 标准化向量
     orthogonal_vector[-200] = 20000
     orthogonal_vector[-100] = 20000
     orthogonal_unit_vector = orthogonal_vector / torch.norm(orthogonal_vector)
     return orthogonal_unit_vector
 
 
 class SignGuardAdversary(Adversary):
-    def __init__(self, clients, global_config: Dict = None):
-        super().__init__(clients, global_config)
-
     def on_local_round_end(self, algorithm: Algorithm):
         updates = self._attack_sign_guard(algorithm)
         for result in algorithm.local_results:

blades/aggregators/aggregators.py

@@ -95,7 +95,7 @@ def obj_func(median, inputs, weights):
             # return (torch.sum(norms * weights) / torch.sum(weights)).item()
 
             return np.average(
-                [torch.norm(p - median).item() for p in inputs],
+                [torch.norm(p - median, p=1).item() for p in inputs],
                 weights=weights.cpu(),
             )
 
@@ -107,7 +107,7 @@ def obj_func(median, inputs, weights):
             # Weiszfeld iterations
             for _ in range(maxiter):
                 prev_obj_value = objective_value
-                denom = torch.stack([torch.norm(p - median) for p in inputs])
+                denom = torch.stack([torch.norm(p - median, p=1) for p in inputs])
                 new_weights = weights / torch.clamp(denom, min=eps)
                 median = weighted_average(inputs, new_weights)
 

blades/algorithms/fedavg/fedavg.py

@@ -112,44 +112,6 @@ def get_default_config(cls) -> FedavgTrainerConfig:
     def setup(self, config: FedavgTrainerConfig):
         super().setup(config)
 
-        # self.adversary = self.config.get_adversary_config().build(
-        #     self.client_manager.clients[: self.config.num_malicious_clients]
-        # )
-        # self.adversary.on_algorithm_start(self)
-
-    # def training_step(self):
-    #     self.worker_group.sync_weights(self.server.get_global_model().state_dict())
-
-    #     def local_training(worker, client):
-    #         dataset = worker.dataset.get_client_dataset(client.client_id)
-    #         result = client.train_one_round(dataset)
-    #         return result
-
-    #     clients = self.client_manager.trainable_clients
-    #     self.local_results = self.worker_group.foreach_execution(
-    #         local_training, clients
-    #     )
-
-    #     self.adversary.on_local_round_end(self)
-    #     updates = [result.pop(CLIENT_UPDATE, None) for result in self.local_results]
-
-    #     losses = []
-    #     for result in self.local_results:
-    #         client = self.client_manager.get_client_by_id(result[CLIENT_ID])
-    #         if not client.is_malicious:
-    #             loss = result.pop("avg_loss")
-    #             losses.append(loss)
-
-    #     self._counters[NUM_GLOBAL_STEPS] += 1
-    #     global_vars = {
-    #         "timestep": self._counters[NUM_GLOBAL_STEPS],
-    #     }
-    #     results = {"train_loss": np.mean(losses)}
-    #     server_return = self.server.step(updates, global_vars)
-    #     results.update(server_return)
-
-    #     return results
-
     def evaluate(self):
         self.worker_group.sync_weights(self.server.get_global_model().state_dict())
 

blades/tuned_examples/fedavg_fashion_mnist.yaml

@@ -13,6 +13,10 @@ fedavg_blades:
       num_clients: 60
       train_batch_size: 64
 
+      partitioner_config:
+        type: IIDPartitioner
+        num_clients: 60
+
     evaluation_interval: 50
 
     num_remote_workers: 0
@@ -21,11 +25,6 @@ fedavg_blades:
     num_cpus_for_driver: 2
     num_gpus_for_driver: 0.3
 
-    # num_remote_workers: 14
-    # num_gpus_per_worker: 0.25
-    # num_cpus_per_worker: 2
-    # num_cpus_for_driver: 2
-    # num_gpus_for_driver: 0.5
 
     # num_clients: 2
     # global_model:
@@ -47,12 +46,12 @@ fedavg_blades:
       aggregator:
         grid_search: [
           # type: Mean,
-          type: Clippedclustering,
-          type: Median,
+          # type: Clippedclustering,
+          # type: Median,
           type: GeoMed,
-          type: DnC,
-          type: Trimmedmean,
-          type: Signguard,
+          # type: DnC,
+          # type: Trimmedmean,
+          # type: Signguard,
           ]
 
       optimizer:
@@ -67,18 +66,18 @@ fedavg_blades:
 
     num_malicious_clients:
       # grid_search: [0]
-      grid_search:  [3, 6, 9, 12, 15, 18]
+      grid_search:  [9, 12, 15]
     adversary_config:
       # type: blades.adversaries.LabelFlipAdversary
       # type: blades.adversaries.SignFlipAdversary
       # type: blades.adversaries.AdaptiveAdversary
       grid_search:
-        - type: blades.adversaries.ALIEAdversary
-        - type: blades.adversaries.LabelFlipAdversary
-        - type: blades.adversaries.NoiseAdversary
-        - type: blades.adversaries.SignFlipAdversary
-        - type: blades.adversaries.IPMAdversary
-          scale: 0.1
-        - type: blades.adversaries.IPMAdversary
-          scale: 100
+        # - type: blades.adversaries.ALIEAdversary
+        # - type: blades.adversaries.LabelFlipAdversary
+        # - type: blades.adversaries.NoiseAdversary
+        # - type: blades.adversaries.SignFlipAdversary
+        # - type: blades.adversaries.IPMAdversary
+        #   scale: 0.1
+        # - type: blades.adversaries.IPMAdversary
+        #   scale: 100
         - type: blades.adversaries.MinMaxAdversary

blades/tuned_examples/fedavg_toy_mnist.yaml

@@ -14,7 +14,7 @@ fedavg_blades:
       train_batch_size: 64
       partitioner_config:
         type: DirichletPartitioner
-        # alpha: 100.
+        alpha: 100.
         num_clients: 4
 
     evaluation_interval: 50

blades/tuned_examples/fedsgd_cifar10.yaml

@@ -15,7 +15,7 @@ fedavg_blades:
 
       partitioner_config:
         type: DirichletPartitioner
-        # alpha: 100.
+        alpha: 100.
         num_clients: 60
 
     evaluation_interval: 50
@@ -41,9 +41,10 @@ fedavg_blades:
     server_config:
       aggregator:
         grid_search: [
-          type: Mean,
+          # type: Mean,
           # type: Clippedclustering,
           # type: Median,
+          type: blades.aggregators.GeoMed,
           # type: GeoMed,
           # type: DnC,
           # type: Trimmedmean,
@@ -68,10 +69,10 @@ fedavg_blades:
       grid_search:
         # - type: blades.adversaries.ALIEAdversary
         # - type: blades.adversaries.LabelFlipAdversary
-        # - type: blades.adversaries.NoiseAdversary
+        - type: blades.adversaries.NoiseAdversary
         - type: blades.adversaries.SignFlipAdversary
         # - type: blades.adversaries.IPMAdversary
         #   scale: 0.1
         # - type: blades.adversaries.IPMAdversary
         #   scale: 100
-        # - type: blades.adversaries.MinMaxAdversary
+        - type: blades.adversaries.MinMaxAdversary

blades/tuned_examples/fedsgd_cnn_fashion_mnist.yaml

@@ -11,6 +11,9 @@ fedavg_blades:
       type: FashionMNIST
       num_clients: 60
       train_batch_size: 64
+      partitioner_config:
+        type: IIDPartitioner
+        num_clients: 60
 
     evaluation_interval: 50
 
@@ -36,12 +39,12 @@ fedavg_blades:
       aggregator:
         grid_search: [
           # type: Mean,
-          type: Clippedclustering,
-          type: Median,
+          # type: Clippedclustering,
+          # type: Median,
           type: GeoMed,
-          type: DnC,
-          type: Trimmedmean,
-          type: Signguard,
+          # type: DnC,
+          # type: Trimmedmean,
+          # type: Signguard,
           # type: Multikrum,
           # type: Centeredclipping
           ]
@@ -56,16 +59,16 @@ fedavg_blades:
     num_malicious_clients:
       # grid_search:  [0,9]
       # grid_search:  [3, 6]
-      grid_search:  [3, 6, 9, 12, 15, 18]
+      grid_search:  [12, 15, 18]
 
     adversary_config:
       grid_search:
         - type: blades.adversaries.ALIEAdversary
         - type: blades.adversaries.LabelFlipAdversary
         - type: blades.adversaries.NoiseAdversary
         - type: blades.adversaries.SignFlipAdversary
-        - type: blades.adversaries.IPMAdversary
-          scale: 0.1
-        - type: blades.adversaries.IPMAdversary
-          scale: 100
+        # - type: blades.adversaries.IPMAdversary
+        #   scale: 0.1
+        # - type: blades.adversaries.IPMAdversary
+        #   scale: 100
         - type: blades.adversaries.MinMaxAdversary

blades/tuned_examples/local20_resnet_cifar10.yaml

@@ -16,9 +16,9 @@ fedavg_blades:
       train_batch_size: 64
 
       partitioner_config:
-        type: DirichletPartitioner
-        alpha:
-          grid_search: [0.1, 1.0]
+        type: IIDPartitioner
+        # alpha:
+          # grid_search: [0.1, 1.0]
         num_clients: 20
 
     evaluation_interval: 50