remove opacus and DP

README.md

@@ -25,7 +25,7 @@ Figure 1: An Example for FedAvg. You can create a scenario using `generate_DATA.
 
 The origin of the **statistical heterogeneity** phenomenon is the personalization of users, who generate non-IID (not Independent and Identically Distributed) and unbalanced data. With statistical heterogeneity existing in the FL scenario, a myriad of approaches have been proposed to crack this hard nut. In contrast, the personalized FL (pFL) may take advantage of the statistically heterogeneous data to learn the personalized model for each user. 
 
-Thanks to [@Stonesjtu](https://github.com/Stonesjtu/pytorch_memlab/blob/d590c489236ee25d157ff60ecd18433e8f9acbe3/pytorch_memlab/mem_reporter.py#L185), this library can also record the **GPU memory usage** for the model. By using the package [opacus](https://opacus.ai/), we introduce **DP (differential privacy)** into this library (please refer to `./system/flcore/clients/clientavg.py` for example). Following [FedCG](https://www.ijcai.org/proceedings/2022/0324.pdf), we also introduce the **[DLG (Deep Leakage from Gradients)](https://papers.nips.cc/paper_files/paper/2019/hash/60a6c4002cc7b29142def8871531281a-Abstract.html) attack** and **PSNR (Peak Signal-to-Noise Ratio) metric** to evaluate the privacy-preserving ability of tFL/pFL algorithms (please refer to `./system/flcore/servers/serveravg.py` for example). *Now we can train on some clients and evaluate performance on other new clients by setting `args.num_new_clients` in `./system/main.py`. Note that not all the tFL/pFL algorithms support this feature.*
+Thanks to [@Stonesjtu](https://github.com/Stonesjtu/pytorch_memlab/blob/d590c489236ee25d157ff60ecd18433e8f9acbe3/pytorch_memlab/mem_reporter.py#L185), this library can also record the **GPU memory usage** for the model. Following [FedCG](https://www.ijcai.org/proceedings/2022/0324.pdf), we also introduce the **[DLG (Deep Leakage from Gradients)](https://papers.nips.cc/paper_files/paper/2019/hash/60a6c4002cc7b29142def8871531281a-Abstract.html) attack** and **PSNR (Peak Signal-to-Noise Ratio) metric** to evaluate the privacy-preserving ability of tFL/pFL algorithms (please refer to `./system/flcore/servers/serveravg.py` for example). *Now we can train on some clients and evaluate performance on other new clients by setting `args.num_new_clients` in `./system/main.py`. Note that not all the tFL/pFL algorithms support this feature.*
 
 **Citation**
 

env_cuda_latest.yaml

@@ -20,7 +20,6 @@ dependencies:
     - torchvision
     - calmsize
     - memory-profiler
-    - opacus
     - portalocker
     - cvxpy
     - higher

system/flcore/clients/clientavg.py

@@ -20,7 +20,6 @@
 import numpy as np
 import time
 from flcore.clients.clientbase import Client
-from utils.privacy import *
 
 
 class clientAVG(Client):
@@ -31,12 +30,6 @@ def train(self):
         trainloader = self.load_train_data()
         # self.model.to(self.device)
         self.model.train()
-
-        # differential privacy
-        if self.privacy:
-            model_origin = copy.deepcopy(self.model)
-            self.model, self.optimizer, trainloader, privacy_engine = \
-                initialize_dp(self.model, self.optimizer, trainloader, self.dp_sigma)
 
         start_time = time.time()
 
@@ -66,12 +59,3 @@ def train(self):
 
         self.train_time_cost['num_rounds'] += 1
         self.train_time_cost['total_cost'] += time.time() - start_time
-
-        if self.privacy:
-            eps, DELTA = get_dp_params(privacy_engine)
-            print(f"Client {self.id}", f"epsilon = {eps:.2f}, sigma = {DELTA}")
-
-            for param, param_dp in zip(model_origin.parameters(), self.model.parameters()):
-                param.data = param_dp.data.clone()
-            self.model = model_origin
-            self.optimizer = torch.optim.SGD(self.model.parameters(), lr=self.learning_rate)

system/flcore/clients/clientbase.py

@@ -59,9 +59,6 @@ def __init__(self, args, id, train_samples, test_samples, **kwargs):
         self.train_time_cost = {'num_rounds': 0, 'total_cost': 0.0}
         self.send_time_cost = {'num_rounds': 0, 'total_cost': 0.0}
 
-        self.privacy = args.privacy
-        self.dp_sigma = args.dp_sigma
-
         self.loss = nn.CrossEntropyLoss()
         self.optimizer = torch.optim.SGD(self.model.parameters(), lr=self.learning_rate)
         self.learning_rate_scheduler = torch.optim.lr_scheduler.ExponentialLR(

system/flcore/clients/clientntd.py

@@ -22,7 +22,6 @@
 import time
 import torch.nn.functional as F
 from flcore.clients.clientbase import Client
-from utils.privacy import *
 
 
 class clientNTD(Client):
@@ -39,12 +38,6 @@ def train(self):
         trainloader = self.load_train_data()
         # self.model.to(self.device)
         self.model.train()
-
-        # differential privacy
-        if self.privacy:
-            model_origin = copy.deepcopy(self.model)
-            self.model, self.optimizer, trainloader, privacy_engine = \
-                initialize_dp(self.model, self.optimizer, trainloader, self.dp_sigma)
 
         start_time = time.time()
 
@@ -76,15 +69,6 @@ def train(self):
 
         self.train_time_cost['num_rounds'] += 1
         self.train_time_cost['total_cost'] += time.time() - start_time
-
-        if self.privacy:
-            eps, DELTA = get_dp_params(privacy_engine)
-            print(f"Client {self.id}", f"epsilon = {eps:.2f}, sigma = {DELTA}")
-
-            for param, param_dp in zip(model_origin.parameters(), self.model.parameters()):
-                param.data = param_dp.data.clone()
-            self.model = model_origin
-            self.optimizer = torch.optim.SGD(self.model.parameters(), lr=self.learning_rate)
 
     def set_parameters(self, model):
         for new_param, old_param in zip(model.parameters(), self.model.parameters()):

system/main.py

@@ -413,9 +413,6 @@ def run(args):
                         help="Running times")
     parser.add_argument('-eg', "--eval_gap", type=int, default=1,
                         help="Rounds gap for evaluation")
-    parser.add_argument('-dp', "--privacy", type=bool, default=False,
-                        help="differential privacy")
-    parser.add_argument('-dps', "--dp_sigma", type=float, default=0.0)
     parser.add_argument('-sfn', "--save_folder_name", type=str, default='items')
     parser.add_argument('-ab', "--auto_break", type=bool, default=False)
     parser.add_argument('-dlg', "--dlg_eval", type=bool, default=False)
@@ -515,9 +512,6 @@ def run(args):
     print("Number of classes: {}".format(args.num_classes))
     print("Backbone: {}".format(args.model))
     print("Using device: {}".format(args.device))
-    print("Using DP: {}".format(args.privacy))
-    if args.privacy:
-        print("Sigma for DP: {}".format(args.dp_sigma))
     print("Auto break: {}".format(args.auto_break))
     if not args.auto_break:
         print("Global rounds: {}".format(args.global_rounds))