add several Byzantine robust algorithms (#552)

LICENSE

@@ -642,3 +642,22 @@ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 SOFTWARE.
+
+---------------------------------------------------------------------------------
+The implementations of median aggregator in federatedscope/core/aggregators/median_aggregator.py
+and trimmedmean aggregator in federatedscope/core/aggregators/trimmedmean_aggregator.py
+are adapted from https://github.com/bladesteam/blades (Apache License)
+
+Copyright (c) 2022 lishenghui
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.

federatedscope/core/aggregators/__init__.py

@@ -8,6 +8,13 @@
     import ServerClientsInterpolateAggregator
 from federatedscope.core.aggregators.fedopt_aggregator import FedOptAggregator
 from federatedscope.core.aggregators.krum_aggregator import KrumAggregator
+from federatedscope.core.aggregators.median_aggregator import MedianAggregator
+from federatedscope.core.aggregators.trimmedmean_aggregator import \
+    TrimmedmeanAggregator
+from federatedscope.core.aggregators.bulyan_aggregator import \
+    BulyanAggregator
+from federatedscope.core.aggregators.normbounding_aggregator import \
+    NormboundingAggregator
 
 __all__ = [
     'Aggregator',
@@ -18,4 +25,8 @@
     'ServerClientsInterpolateAggregator',
     'FedOptAggregator',
     'KrumAggregator',
+    'MedianAggregator',
+    'TrimmedmeanAggregator',
+    'BulyanAggregator',
+    'NormboundingAggregator',
 ]

federatedscope/core/aggregators/bulyan_aggregator.py

@@ -0,0 +1,106 @@
+import copy
+import torch
+from federatedscope.core.aggregators import ClientsAvgAggregator
+
+
+class BulyanAggregator(ClientsAvgAggregator):
+    """
+    Implementation of Bulyan refers to `The Hidden Vulnerability
+    of Distributed Learning in Byzantium`
+    [Mhamdi et al., 2018]
+    (http://proceedings.mlr.press/v80/mhamdi18a/mhamdi18a.pdf)
+
+    It combines the MultiKrum aggregator and the treamedmean aggregator
+    """
+    def __init__(self, model=None, device='cpu', config=None):
+        super(BulyanAggregator, self).__init__(model, device, config)
+        self.byzantine_node_num = config.aggregator.byzantine_node_num
+        self.sample_client_rate = config.federate.sample_client_rate
+        assert 4 * self.byzantine_node_num + 3 <= config.federate.client_num
+
+    def aggregate(self, agg_info):
+        """
+        To preform aggregation with Median aggregation rule
+        Arguments:
+        agg_info (dict): the feedbacks from clients
+        :returns: the aggregated results
+        :rtype: dict
+        """
+        models = agg_info["client_feedback"]
+        avg_model = self._aggre_with_bulyan(models)
+        updated_model = copy.deepcopy(avg_model)
+        init_model = self.model.state_dict()
+        for key in avg_model:
+            updated_model[key] = init_model[key] + avg_model[key]
+        return updated_model
+
+    def _calculate_distance(self, model_a, model_b):
+        """
+        Calculate the Euclidean distance between two given model para delta
+        """
+        distance = 0.0
+
+        for key in model_a:
+            if isinstance(model_a[key], torch.Tensor):
+                model_a[key] = model_a[key].float()
+                model_b[key] = model_b[key].float()
+            else:
+                model_a[key] = torch.FloatTensor(model_a[key])
+                model_b[key] = torch.FloatTensor(model_b[key])
+
+            distance += torch.dist(model_a[key], model_b[key], p=2)
+        return distance
+
+    def _calculate_score(self, models):
+        """
+        Calculate Krum scores
+        """
+        model_num = len(models)
+        closest_num = model_num - self.byzantine_node_num - 2
+
+        distance_matrix = torch.zeros(model_num, model_num)
+        for index_a in range(model_num):
+            for index_b in range(index_a, model_num):
+                if index_a == index_b:
+                    distance_matrix[index_a, index_b] = float('inf')
+                else:
+                    distance_matrix[index_a, index_b] = distance_matrix[
+                        index_b, index_a] = self._calculate_distance(
+                            models[index_a], models[index_b])
+
+        sorted_distance = torch.sort(distance_matrix)[0]
+        krum_scores = torch.sum(sorted_distance[:, :closest_num], axis=-1)
+        return krum_scores
+
+    def _aggre_with_bulyan(self, models):
+        '''
+        Apply MultiKrum to select \theta (\theta <= client_num-
+        2*self.byzantine_node_num) local models
+        '''
+        init_model = self.model.state_dict()
+        global_update = copy.deepcopy(init_model)
+        models_para = [each_model[1] for each_model in models]
+        krum_scores = self._calculate_score(models_para)
+        index_order = torch.sort(krum_scores)[1].numpy()
+        reliable_models = list()
+        for number, index in enumerate(index_order):
+            if number < len(models) - int(
+                    2 * self.sample_client_rate * self.byzantine_node_num):
+                reliable_models.append(models[index])
+        '''
+        Sort parameter for each coordinate of the rest \theta reliable
+        local models, and find \gamma (gamma<\theta-2*self.byzantine_num)
+        parameters closest to the median to perform averaging
+        '''
+        exluded_num = int(self.sample_client_rate * self.byzantine_node_num)
+        gamma = len(reliable_models) - 2 * exluded_num
+        for key in init_model:
+            temp = torch.stack(
+                [each_model[1][key] for each_model in reliable_models], 0)
+            pos_largest, _ = torch.topk(temp, exluded_num, 0)
+            neg_smallest, _ = torch.topk(-temp, exluded_num, 0)
+            new_stacked = torch.cat([temp, -pos_largest,
+                                     neg_smallest]).sum(0).float()
+            new_stacked /= gamma
+            global_update[key] = new_stacked
+        return global_update

federatedscope/core/aggregators/krum_aggregator.py

@@ -14,7 +14,7 @@ class KrumAggregator(ClientsAvgAggregator):
     def __init__(self, model=None, device='cpu', config=None):
         super(KrumAggregator, self).__init__(model, device, config)
         self.byzantine_node_num = config.aggregator.byzantine_node_num
-        self.krum_agg_num = config.aggregator.krum.agg_num
+        self.krum_agg_num = config.aggregator.BFT_args.krum_agg_num
         assert 2 * self.byzantine_node_num + 2 < config.federate.client_num, \
             "it should be satisfied that 2*byzantine_node_num + 2 < client_num"
 

federatedscope/core/aggregators/median_aggregator.py

@@ -0,0 +1,52 @@
+import copy
+import torch
+import numpy as np
+from federatedscope.core.aggregators import ClientsAvgAggregator
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+class MedianAggregator(ClientsAvgAggregator):
+    """
+    Implementation of median refers to `Byzantine-robust distributed
+    learning: Towards optimal statistical rates`
+    [Yin et al., 2018]
+    (http://proceedings.mlr.press/v80/yin18a/yin18a.pdf)
+
+    It computes the coordinate-wise median of recieved updates from clients
+
+    The code is adapted from https://github.com/bladesteam/blades
+    """
+    def __init__(self, model=None, device='cpu', config=None):
+        super(MedianAggregator, self).__init__(model, device, config)
+        self.byzantine_node_num = config.aggregator.byzantine_node_num
+        assert 2 * self.byzantine_node_num + 2 < config.federate.client_num, \
+            "it should be satisfied that 2*byzantine_node_num + 2 < client_num"
+
+    def aggregate(self, agg_info):
+        """
+        To preform aggregation with Median aggregation rule
+        Arguments:
+        agg_info (dict): the feedbacks from clients
+        :returns: the aggregated results
+        :rtype: dict
+        """
+        models = agg_info["client_feedback"]
+        avg_model = self._aggre_with_median(models)
+        updated_model = copy.deepcopy(avg_model)
+        init_model = self.model.state_dict()
+        for key in avg_model:
+            updated_model[key] = init_model[key] + avg_model[key]
+        return updated_model
+
+    def _aggre_with_median(self, models):
+        init_model = self.model.state_dict()
+        global_update = copy.deepcopy(init_model)
+        for key in init_model:
+            temp = torch.stack([each_model[1][key] for each_model in models],
+                               0)
+            temp_pos, _ = torch.median(temp, dim=0)
+            temp_neg, _ = torch.median(-temp, dim=0)
+            global_update[key] = (temp_pos - temp_neg) / 2
+        return global_update

federatedscope/core/aggregators/normbounding_aggregator.py

@@ -0,0 +1,64 @@
+import logging
+import copy
+import torch
+import numpy as np
+from federatedscope.core.aggregators import ClientsAvgAggregator
+
+logger = logging.getLogger(__name__)
+
+
+class NormboundingAggregator(ClientsAvgAggregator):
+    """
+    The server clips each update to reduce the negative impact \
+        of malicious updates.
+    """
+    def __init__(self, model=None, device='cpu', config=None):
+        super(NormboundingAggregator, self).__init__(model, device, config)
+        self.norm_bound = config.aggregator.BFT_args.normbounding_norm_bound
+
+    def aggregate(self, agg_info):
+        """
+        To preform aggregation with normbounding aggregation rule
+        Arguments:
+        agg_info (dict): the feedbacks from clients
+        :returns: the aggregated results
+        :rtype: dict
+        """
+        models = agg_info["client_feedback"]
+        avg_model = self._aggre_with_normbounding(models)
+        updated_model = copy.deepcopy(avg_model)
+        init_model = self.model.state_dict()
+        for key in avg_model:
+            updated_model[key] = init_model[key] + avg_model[key]
+        return updated_model
+
+    def _aggre_with_normbounding(self, models):
+        models_temp = []
+        for each_model in models:
+            param = self._flatten_updates(each_model[1])
+            if torch.norm(param, p=2) > self.norm_bound:
+                scaling_rate = self.norm_bound / torch.norm(param, p=2)
+                scaled_param = scaling_rate * param
+                models_temp.append(
+                    (each_model[0], self._reconstruct_updates(scaled_param)))
+            else:
+                models_temp.append(each_model)
+        return self._para_weighted_avg(models_temp)
+
+    def _flatten_updates(self, model):
+        model_update = []
+        init_model = self.model.state_dict()
+        for key in init_model:
+            model_update.append(model[key].view(-1))
+        return torch.cat(model_update, dim=0)
+
+    def _reconstruct_updates(self, flatten_updates):
+        start_idx = 0
+        init_model = self.model.state_dict()
+        reconstructed_model = copy.deepcopy(init_model)
+        for key in init_model:
+            reconstructed_model[key] = flatten_updates[
+                start_idx:start_idx + len(init_model[key].view(-1))].reshape(
+                    init_model[key].shape)
+            start_idx = start_idx + len(init_model[key].view(-1))
+        return reconstructed_model

federatedscope/core/aggregators/trimmedmean_aggregator.py

@@ -0,0 +1,57 @@
+import copy
+import torch
+import numpy as np
+from federatedscope.core.aggregators import ClientsAvgAggregator
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+class TrimmedmeanAggregator(ClientsAvgAggregator):
+    """
+    Implementation of median refer to `Byzantine-robust distributed
+    learning: Towards optimal statistical rates`
+    [Yin et al., 2018]
+    (http://proceedings.mlr.press/v80/yin18a/yin18a.pdf)
+
+    The code is adapted from https://github.com/bladesteam/blades
+    """
+    def __init__(self, model=None, device='cpu', config=None):
+        super(TrimmedmeanAggregator, self).__init__(model, device, config)
+        self.excluded_ratio = \
+            config.aggregator.BFT_args.trimmedmean_excluded_ratio
+        self.byzantine_node_num = config.aggregator.byzantine_node_num
+        assert 2 * self.byzantine_node_num + 2 < config.federate.client_num, \
+            "it should be satisfied that 2*byzantine_node_num + 2 < client_num"
+        assert self.excluded_ratio < 0.5
+
+    def aggregate(self, agg_info):
+        """
+        To preform aggregation with trimmedmean aggregation rule
+        Arguments:
+        agg_info (dict): the feedbacks from clients
+        :returns: the aggregated results
+        :rtype: dict
+        """
+        models = agg_info["client_feedback"]
+        avg_model = self._aggre_with_trimmedmean(models)
+        updated_model = copy.deepcopy(avg_model)
+        init_model = self.model.state_dict()
+        for key in avg_model:
+            updated_model[key] = init_model[key] + avg_model[key]
+        return updated_model
+
+    def _aggre_with_trimmedmean(self, models):
+        init_model = self.model.state_dict()
+        global_update = copy.deepcopy(init_model)
+        excluded_num = int(len(models) * self.excluded_ratio)
+        for key in init_model:
+            temp = torch.stack([each_model[1][key] for each_model in models],
+                               0)
+            pos_largest, _ = torch.topk(temp, excluded_num, 0)
+            neg_smallest, _ = torch.topk(-temp, excluded_num, 0)
+            new_stacked = torch.cat([temp, -pos_largest,
+                                     neg_smallest]).sum(0).float()
+            new_stacked /= len(temp) - 2 * excluded_num
+            global_update[key] = new_stacked
+        return global_update

federatedscope/core/auxiliaries/aggregator_builder.py

@@ -58,7 +58,18 @@ def get_aggregator(method, model=None, device=None, online=False, config=None):
         from federatedscope.core.aggregators import ClientsAvgAggregator, \
             OnlineClientsAvgAggregator, ServerClientsInterpolateAggregator, \
             FedOptAggregator, NoCommunicationAggregator, \
-            AsynClientsAvgAggregator, KrumAggregator
+            AsynClientsAvgAggregator, KrumAggregator, \
+            MedianAggregator, TrimmedmeanAggregator, \
+            BulyanAggregator,  NormboundingAggregator
+
+    STR2AGG = {
+        'fedavg': ClientsAvgAggregator,
+        'krum': KrumAggregator,
+        'median': MedianAggregator,
+        'bulyan': BulyanAggregator,
+        'trimmedmean': TrimmedmeanAggregator,
+        'normbounding': NormboundingAggregator
+    }
 
     if method.lower() in constants.AGGREGATOR_TYPE:
         aggregator_type = constants.AGGREGATOR_TYPE[method.lower()]
@@ -87,12 +98,17 @@ def get_aggregator(method, model=None, device=None, online=False, config=None):
             return AsynClientsAvgAggregator(model=model,
                                             device=device,
                                             config=config)
-        elif config.aggregator.krum.use:
-            return KrumAggregator(model=model, device=device, config=config)
         else:
-            return ClientsAvgAggregator(model=model,
-                                        device=device,
-                                        config=config)
+            if config.aggregator.robust_rule not in STR2AGG:
+                logger.warning(
+                    f'The specified {config.aggregator.robust_rule} aggregtion\
+                    rule has not been supported, the vanilla fedavg algorithm \
+                    will be used instead.')
+            return STR2AGG.get(config.aggregator.robust_rule,
+                               ClientsAvgAggregator)(model=model,
+                                                     device=device,
+                                                     config=config)
+
     elif aggregator_type == 'server_clients_interpolation':
         return ServerClientsInterpolateAggregator(
             model=model,