-
Notifications
You must be signed in to change notification settings - Fork 56
/
sync_dp_servers.py
192 lines (161 loc) · 7.01 KB
/
sync_dp_servers.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
#!/usr/bin/env python3
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.
# pyre-unsafe
from __future__ import annotations
from dataclasses import dataclass
from typing import Optional
from flsim.channels.base_channel import IdentityChannel, IFLChannel
from flsim.channels.message import Message
from flsim.data.data_provider import IFLDataProvider
from flsim.interfaces.model import IFLModel
from flsim.optimizers.server_optimizers import FedAvgOptimizerConfig
from flsim.privacy.common import PrivacyBudget, PrivacySetting
from flsim.privacy.privacy_engine import IPrivacyEngine
from flsim.privacy.privacy_engine_factory import NoiseType, PrivacyEngineFactory
from flsim.privacy.user_update_clip import IUserClipper
from flsim.servers.aggregator import AggregationType, Aggregator
from flsim.servers.sync_servers import ISyncServer, SyncServerConfig
from flsim.utils.config_utils import fullclassname, init_self_cfg
from flsim.utils.distributed.fl_distributed import FLDistributedUtils, OperationType
from flsim.utils.fl.common import FLModelParamUtils
from hydra.utils import instantiate
from omegaconf import OmegaConf
from papaya.toolkit.simulation.flsim.active_user_selectors.simple_user_selector import (
UniformlyRandomActiveUserSelectorConfig,
)
class SyncDPSGDServer(ISyncServer):
"""
User level DP-SGD Server implementing https://arxiv.org/abs/1710.06963
Args:
global_model: IFLModel: Global (server model) to be updated between rounds
users_per_round: int: User per round to calculate sampling rate
num_total_users: int: Total users in the dataset to calculate sampling rate
Sampling rate = users_per_round / num_total_users
channel: Optional[IFLChannel]: Communication channel between server and clients
"""
def __init__(
self,
*,
global_model: IFLModel,
channel: Optional[IFLChannel] = None,
**kwargs,
):
init_self_cfg(
self,
component_class=__class__, # pyre-fixme[10]: Name `__class__` is used but not defined.
config_class=SyncDPSGDServerConfig,
**kwargs,
)
assert (
self.cfg.aggregation_type == AggregationType.AVERAGE # pyre-ignore[16]
), "DP training must be done with simple averaging and uniform weights."
self.privacy_budget = PrivacyBudget()
self._optimizer = instantiate(
config=self.cfg.server_optimizer,
model=global_model.fl_get_module(),
)
self._global_model: IFLModel = global_model
self._clipping_value = self.cfg.privacy_setting.clipping.clipping_value
self._user_update_clipper = IUserClipper.create_clipper(
self.cfg.privacy_setting
)
self._aggregator: Aggregator = Aggregator(
module=global_model.fl_get_module(),
aggregation_type=self.cfg.aggregation_type,
only_federated_params=self.cfg.only_federated_params,
)
self._privacy_engine: Optional[IPrivacyEngine] = None
self._active_user_selector = instantiate(self.cfg.active_user_selector)
self._channel: IFLChannel = channel or IdentityChannel()
@classmethod
def _set_defaults_in_cfg(cls, cfg):
if OmegaConf.is_missing(cfg.active_user_selector, "_target_"):
cfg.active_user_selector = UniformlyRandomActiveUserSelectorConfig()
if OmegaConf.is_missing(cfg.server_optimizer, "_target_"):
cfg.server_optimizer = FedAvgOptimizerConfig()
@property
def global_model(self):
return self._global_model
def select_clients_for_training(
self,
num_total_users,
users_per_round,
data_provider: Optional[IFLDataProvider] = None,
global_round_num: Optional[int] = None,
):
if self._privacy_engine is None:
self._privacy_engine: IPrivacyEngine = PrivacyEngineFactory.create(
# pyre-ignore[16]
self.cfg.privacy_setting,
users_per_round,
num_total_users,
noise_type=NoiseType.GAUSSIAN,
)
self._privacy_engine.attach(self._global_model.fl_get_module())
return self._active_user_selector.get_user_indices(
num_total_users=num_total_users,
users_per_round=users_per_round,
data_provider=data_provider,
global_round_num=global_round_num,
)
def init_round(self):
self._aggregator.zero_weights()
self._optimizer.zero_grad()
self._privacy_engine.attach(self._global_model.fl_get_module())
self._user_update_clipper.reset_clipper_stats()
def receive_update_from_client(self, message: Message):
message = self._channel.client_to_server(message)
self._aggregator.apply_weight_to_update(
delta=message.model.fl_get_module(), weight=message.weight
)
self._user_update_clipper.clip(message.model.fl_get_module())
self._aggregator.add_update(
delta=message.model.fl_get_module(), weight=message.weight
)
def mock_step(self):
"""
Populates the grad attributes of the global model without performing
an optimization step and returns the aggregated *sum* and not avearge.
Useful for the CANIFE Empirical Privacy Measurement method.
"""
# mock step that updates the gradients of the global model
self.step(mock=True)
# scale by sum_weights
model = self._global_model.fl_get_module()
FLModelParamUtils.multiply_gradient_by_weight(
model,
self._aggregator.sum_weights.item(),
model,
)
return model
def step(self, mock=False):
assert self._privacy_engine is not None, "PrivacyEngine is not initialized"
aggregated_model = self._aggregator.aggregate(distributed_op=OperationType.SUM)
if FLDistributedUtils.is_master_worker():
self._privacy_engine.add_noise(
aggregated_model,
self._user_update_clipper.max_norm
/ self._aggregator.sum_weights.item(),
)
FLDistributedUtils.synchronize_model_across_workers(
operation=OperationType.BROADCAST,
model=aggregated_model,
weights=self._aggregator.sum_weights,
)
FLModelParamUtils.set_gradient(
model=self._global_model.fl_get_module(),
reference_gradient=aggregated_model,
)
if not mock:
self._optimizer.step()
self.privacy_budget = self._privacy_engine.get_privacy_spent()
self._user_update_clipper.update_clipper_stats()
@dataclass
class SyncDPSGDServerConfig(SyncServerConfig):
_target_: str = fullclassname(SyncDPSGDServer)
aggregation_type: AggregationType = AggregationType.AVERAGE
privacy_setting: PrivacySetting = PrivacySetting()