for idx in idxs_users:
local_model = LocalUpdate(args=args,
dataset=train_dataset,
idxs=user_groups[idx],
logger=logger)
w, loss = local_model.update_weights(
model=copy.deepcopy(global_model), global_round=epoch)
local_weights.append(copy.deepcopy(w))
local_losses.append(copy.deepcopy(loss))
# update global weights
global_weights = average_weights(local_weights)
# update global weights
global_model.load_state_dict(global_weights)
loss_avg = sum(local_losses) / len(local_losses)
train_loss.append(loss_avg)
# Calculate avg training accuracy over all users at every epoch
list_acc, list_loss = [], []
global_model.eval()
for c in range(args.num_users):
local_model = LocalUpdate(args=args,
dataset=train_dataset,
idxs=user_groups[idx],
logger=logger)
acc, loss = local_model.inference(model=global_model)
list_acc.append(acc)
list_loss.append(loss)
def poisoned_pixel_LDP(norm_bound, noise_scale, nb_attackers, seed=1):
start_time = time.time()
# define paths
path_project = os.path.abspath('..')
logger = SummaryWriter('../logs')
args = args_parser()
exp_details(args)
# set seed
torch.manual_seed(seed)
np.random.seed(seed)
# device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# load dataset and user groups
train_dataset, test_dataset, user_groups = get_dataset(args)
# BUILD MODEL
if args.model == 'cnn':
# Convolutional neural netork
if args.dataset == 'mnist':
global_model = CNNMnist(args=args)
elif args.dataset == 'fmnist':
global_model = CNNFashion_Mnist(args=args)
elif args.dataset == 'cifar':
global_model = CNNCifar(args=args)
elif args.model == 'mlp':
# Multi-layer preceptron
img_size = train_dataset[0][0].shape
len_in = 1
for x in img_size:
len_in *= x
global_model = MLP(dim_in=len_in,
dim_hidden=64,
dim_out=args.num_classes)
else:
exit('Error: unrecognized model')
# Set the model to train and send it to device.
global_model.to(device)
global_model.train()
print(global_model)
# copy weights
global_weights = global_model.state_dict()
# testing accuracy for global model
testing_accuracy = [0.1]
backdoor_accuracy = [0.1]
for epoch in tqdm(range(args.epochs)):
local_w, local_norms = [], []
print(f'\n | Global Training Round : {epoch+1} |\n')
global_model.train()
m = max(int(args.frac * args.num_users), 1)
idxs_users = np.random.choice(range(args.num_users), m, replace=False)
# Poisonous updates
for idx in idxs_users[0:nb_attackers]:
local_model = LocalUpdate(args=args,
dataset=train_dataset,
idxs=user_groups[idx],
logger=logger)
w, _ = local_model.pixel_ldp(model=copy.deepcopy(global_model),
norm_bound=norm_bound,
noise_scale=noise_scale)
local_w.append(copy.deepcopy(w))
# Regular updates
for idx in idxs_users[nb_attackers:]:
local_model = LocalUpdate(args=args,
dataset=train_dataset,
idxs=user_groups[idx],
logger=logger)
w, _ = local_model.dp_sgd(model=copy.deepcopy(global_model),
norm_bound=norm_bound,
noise_scale=noise_scale)
local_w.append(copy.deepcopy(w))
# update global weights
global_weights = average_weights(local_w)
global_model.load_state_dict(global_weights)
# test accuracy
test_acc, test_loss, backdoor = test_backdoor_pixel(
args, global_model, test_dataset)
testing_accuracy.append(test_acc)
backdoor_accuracy.append(backdoor)
print("Testing & Backdoor accuracies")
print(testing_accuracy)
print(backdoor_accuracy)
# save test accuracy
np.savetxt(
'../save/PixelAttack/TestAcc/LDP_iid_{}_{}_clip{}_scale{}_attackers{}_seed{}.txt'
.format(args.dataset, args.model, norm_bound, noise_scale,
nb_attackers, s), testing_accuracy)
np.savetxt(
'../save/PixelAttack/BackdoorAcc/LDP_iid_{}_{}_clip{}_scale{}_attackers{}_seed{}.txt'
.format(args.dataset, args.model, norm_bound, noise_scale,
nb_attackers, s), backdoor_accuracy)
def poisoned_pixel_CDP(norm_bound, noise_scale, nb_attackers, seed=1):
start_time = time.time()
# define paths
path_project = os.path.abspath('..')
logger = SummaryWriter('../logs')
args = args_parser()
exp_details(args)
# set seed
torch.manual_seed(seed)
np.random.seed(seed)
# device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# load dataset and user groups
train_dataset, test_dataset, user_groups = get_dataset(args)
# BUILD MODEL
if args.model == 'cnn':
# Convolutional neural netork
if args.dataset == 'mnist':
global_model = CNNMnist(args=args)
elif args.dataset == 'fmnist':
global_model = CNNFashion_Mnist(args=args)
elif args.dataset == 'cifar':
global_model = CNNCifar(args=args)
elif args.model == 'mlp':
# Multi-layer preceptron
img_size = train_dataset[0][0].shape
len_in = 1
for x in img_size:
len_in *= x
global_model = MLP(dim_in=len_in,
dim_hidden=64,
dim_out=args.num_classes)
else:
exit('Error: unrecognized model')
# Set the model to train and send it to device.
global_model.to(device)
global_model.train()
print(global_model)
# copy weights
global_weights = global_model.state_dict()
# load poisoned model
backdoor_model = copy.deepcopy(global_model)
backdoor_model.load_state_dict(torch.load('../save/poison_model.pth'))
# testing accuracy for global model
testing_accuracy = [0.1]
backdoor_accuracy = [0.1]
for epoch in tqdm(range(args.epochs)):
local_del_w, local_norms = [], []
print(f'\n | Global Training Round : {epoch + 1} |\n')
global_model.train()
m = max(int(args.frac * args.num_users), 1)
idxs_users = np.random.choice(range(args.num_users), m, replace=False)
# Adversary updates
print("Evil")
for idx in idxs_users[0:nb_attackers]:
# backdoor model
w = copy.deepcopy(backdoor_model)
# compute change in parameters and norm
zeta = 0
for del_w, w_old in zip(w.parameters(), global_model.parameters()):
del_w.data = del_w.data - copy.deepcopy(w_old.data)
zeta += del_w.norm(2).item()**2
zeta = zeta**(1. / 2)
del_w = w.state_dict()
print("EVIL")
print(zeta)
# add to global round
local_del_w.append(copy.deepcopy(del_w))
local_norms.append(copy.deepcopy(zeta))
# Non-adversarial updates
for idx in idxs_users[nb_attackers:]:
local_model = LocalUpdate(args=args,
dataset=train_dataset,
idxs=user_groups[idx],
logger=logger)
del_w, zeta = local_model.update_weights(
model=copy.deepcopy(global_model), change=1)
local_del_w.append(copy.deepcopy(del_w))
local_norms.append(copy.deepcopy(zeta))
print("good")
#print(zeta)
# norm bound (e.g. median of norms)
clip_factor = norm_bound #min(norm_bound, np.median(local_norms))
print(clip_factor)
# clip updates
for i in range(len(idxs_users)):
for param in local_del_w[i].values():
print(max(1, local_norms[i] / clip_factor))
param /= max(1, local_norms[i] / clip_factor)
# average local model updates
average_del_w = average_weights(local_del_w)
# Update model and add noise
# w_{t+1} = w_{t} + avg(del_w1 + del_w2 + ... + del_wc) + Noise
for param, param_del_w in zip(global_weights.values(),
average_del_w.values()):
param += param_del_w
param += torch.randn(
param.size()) * noise_scale * norm_bound / len(idxs_users)
global_model.load_state_dict(global_weights)
# test accuracy
test_acc, test_loss, backdoor = test_backdoor_pixel(
args, global_model, test_dataset)
testing_accuracy.append(test_acc)
backdoor_accuracy.append(backdoor)
print("Testing & Backdoor accuracies")
print(testing_accuracy)
print(backdoor_accuracy)
# save test accuracy
np.savetxt(
'../save/PixelAttack/TestAcc/iid_GDP_{}_{}_clip{}_scale{}_attackers{}_seed{}.txt'
.format(args.dataset, args.model, norm_bound, noise_scale,
nb_attackers, s), testing_accuracy)
np.savetxt(
'../save/PixelAttack/BackdoorAcc/iid_GDP_{}_{}_clip{}_scale{}_attackers{}_seed{}.txt'
.format(args.dataset, args.model, norm_bound, noise_scale,
nb_attackers, s), backdoor_accuracy)
def poisoned_1to7_NoDefense(seed=1):
start_time = time.time()
# define paths
path_project = os.path.abspath('..')
logger = SummaryWriter('../logs')
args = args_parser()
exp_details(args)
# set seed
torch.manual_seed(seed)
np.random.seed(seed)
# device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# load dataset and user groups
train_dataset, test_dataset, user_groups = get_dataset(args)
# BUILD MODEL
if args.model == 'cnn':
# Convolutional neural netork
if args.dataset == 'mnist':
global_model = CNNMnist(args=args)
elif args.dataset == 'fmnist':
global_model = CNNFashion_Mnist(args=args)
elif args.dataset == 'cifar':
global_model = CNNCifar(args=args)
elif args.model == 'mlp':
# Multi-layer preceptron
img_size = train_dataset[0][0].shape
len_in = 1
for x in img_size:
len_in *= x
global_model = MLP(dim_in=len_in,
dim_hidden=64,
dim_out=args.num_classes)
else:
exit('Error: unrecognized model')
# Set the model to train and send it to device.
global_model.to(device)
global_model.train()
print(global_model)
# copy weights
global_weights = global_model.state_dict()
# testing accuracy for global model
testing_accuracy = [0.1]
backdoor_accuracy = [0.1]
for epoch in tqdm(range(args.epochs)):
local_del_w, local_norms = [], []
print(f'\n | Global Training Round : {epoch+1} |\n')
global_model.train()
m = max(int(args.frac * args.num_users), 1)
idxs_users = np.random.choice(range(args.num_users), m, replace=False)
# Adversary updates
print("Evil norms:")
for idx in idxs_users[0:args.nb_attackers]:
local_model = LocalUpdate(args=args,
dataset=train_dataset,
idxs=user_groups[idx],
logger=logger)
del_w, zeta = local_model.poisoned_1to7(
model=copy.deepcopy(global_model), change=1)
local_del_w.append(copy.deepcopy(del_w))
local_norms.append(copy.deepcopy(zeta))
print(zeta)
# Non-adversarial updates
print("Good norms:")
for idx in idxs_users[args.nb_attackers:]:
print(idx)
local_model = LocalUpdate(args=args,
dataset=train_dataset,
idxs=user_groups[idx],
logger=logger)
del_w, zeta = local_model.update_weights(
model=copy.deepcopy(global_model), change=1)
local_del_w.append(copy.deepcopy(del_w))
local_norms.append(copy.deepcopy(zeta))
print(zeta)
# average local updates
average_del_w = average_weights(local_del_w)
# Update global model: w_{t+1} = w_{t} + average_del_w
for param, param_del_w in zip(global_weights.values(),
average_del_w.values()):
param += param_del_w
global_model.load_state_dict(global_weights)
# test accuracy, backdoor accuracy
test_acc, test_loss, back_acc = test_inference1to7(
args, global_model, test_dataset)
testing_accuracy.append(test_acc)
backdoor_accuracy.append(back_acc)
print("Test & Backdoor accuracy")
print(testing_accuracy)
print(backdoor_accuracy)
# save accuracy
np.savetxt(
'../save/1to7Attack/TestAcc/NoDefense_{}_{}_seed{}.txt'.format(
args.dataset, args.model, s), testing_accuracy)
np.savetxt(
'../save/1to7Attack/BackAcc/NoDefense_{}_{}_seed{}.txt'.format(
args.dataset, args.model, s), backdoor_accuracy)
elif args.dataset == 'cifar100' or args.dataset == 'cifar10':
if args.mode == 'model_heter':
if i < 10:
args.stride = [1, 4]
else:
args.stride = [2, 2]
else:
args.stride = [2, 2]
resnet = resnet18(args,
pretrained=False,
num_classes=args.num_classes)
initial_weight = model_zoo.load_url(model_urls['resnet18'])
local_model = resnet
initial_weight_1 = local_model.state_dict()
for key in initial_weight.keys():
if key[0:3] == 'fc.' or key[0:5] == 'conv1' or key[
0:3] == 'bn1':
initial_weight[key] = initial_weight_1[key]
local_model.load_state_dict(initial_weight)
local_model.to(args.device)
local_model.train()
local_model_list.append(local_model)
if args.mode == 'task_heter':
FedProto_taskheter(args, train_dataset, test_dataset, user_groups,
user_groups_lt, local_model_list, classes_list)
else:
FedProto_modelheter(args, train_dataset, test_dataset, user_groups,
user_groups_lt, local_model_list, classes_list)
local_model = LocalUpdate(
args=args,
dataset=train_dataset,
idxs=user_groups[idx],
logger=logger) # 注意这里发给client的是train_dataset
w, loss = local_model.update_weights(
model=copy.deepcopy(global_model), global_round=epoch)
local_weights.append(
copy.deepcopy(w)) # 记录下client model的模型参数和训练集Loss
local_losses.append(copy.deepcopy(loss))
# update global weights
global_weights = average_weights(local_weights)
# update global weights
global_model.load_state_dict(global_weights) # 将平均后的模型赋值给global model
loss_avg = sum(local_losses) / len(local_losses)
train_loss.append(loss_avg)
# Calculate avg training accuracy over all users at every epoch
list_acc, list_loss = [], []
global_model.eval()
for c in range(args.num_users):
local_model = LocalUpdate(args=args,
dataset=train_dataset,
idxs=user_groups[idx],
logger=logger)
acc, loss = local_model.inference(model=global_model)
list_acc.append(acc)
list_loss.append(loss)
def poisoned_random_CDP(seed=1):
# Central DP to protect against attackers
start_time = time.time()
# define paths
path_project = os.path.abspath('..')
logger = SummaryWriter('../logs')
args = args_parser()
exp_details(args)
# set seed
torch.manual_seed(seed)
np.random.seed(seed)
# device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# load dataset and user groups
train_dataset, test_dataset, user_groups = get_dataset(args)
# BUILD MODEL
if args.model == 'cnn':
# Convolutional neural netork
if args.dataset == 'mnist':
global_model = CNNMnist(args=args)
elif args.dataset == 'fmnist':
global_model = CNNFashion_Mnist(args=args)
elif args.dataset == 'cifar':
global_model = CNNCifar(args=args)
elif args.model == 'mlp':
# Multi-layer preceptron
img_size = train_dataset[0][0].shape
len_in = 1
for x in img_size:
len_in *= x
global_model = MLP(dim_in=len_in,
dim_hidden=64,
dim_out=args.num_classes)
else:
exit('Error: unrecognized model')
# Set the model to train and send it to device.
global_model.to(device)
global_model.train()
print(global_model)
# copy weights
global_weights = global_model.state_dict()
# testing accuracy for global model
testing_accuracy = [0.1]
backdoor_accuracy = [0.1]
for epoch in tqdm(range(args.epochs)):
local_del_w, local_norms = [], []
print(f'\n | Global Training Round : {epoch+1} |\n')
global_model.train()
m = max(int(args.frac * args.num_users), 1)
idxs_users = np.random.choice(range(args.num_users), m, replace=False)
# Adversaries' update
for idx in idxs_users[0:args.nb_attackers]:
local_model = LocalUpdate(args=args,
dataset=train_dataset,
idxs=user_groups[idx],
logger=logger)
del_w, zeta = local_model.poisoned_1to7(
model=copy.deepcopy(global_model), change=1)
local_del_w.append(copy.deepcopy(del_w))
local_norms.append(copy.deepcopy(zeta))
# Non-adversary updates
for idx in idxs_users[args.nb_attackers:]:
local_model = LocalUpdate(args=args,
dataset=train_dataset,
idxs=user_groups[idx],
logger=logger)
del_w, zeta = local_model.update_weights(
model=copy.deepcopy(global_model), change=1)
local_del_w.append(copy.deepcopy(del_w))
local_norms.append(copy.deepcopy(zeta))
# norm bound (e.g. median of norms)
median_norms = args.norm_bound #np.median(local_norms) #args.norm_bound
print(median_norms)
# clip weight updates
for i in range(len(idxs_users)):
for param in local_del_w[i].values():
param /= max(1, local_norms[i] / median_norms)
# average the clipped weight updates
average_del_w = average_weights(local_del_w)
# Update global model using clipped weight updates, and add noise
# w_{t+1} = w_{t} + avg(del_w1 + del_w2 + ... + del_wc) + Noise
for param, param_del_w in zip(global_weights.values(),
average_del_w.values()):
param += param_del_w
param += torch.randn(
param.size()) * args.noise_scale * median_norms / (
len(idxs_users)**0.5)
global_model.load_state_dict(global_weights)
# test accuracy
test_acc, test_loss = test_inference(args, global_model, test_dataset)
testing_accuracy.append(test_acc)
print("Test accuracy")
print(testing_accuracy)
# save test accuracy
np.savetxt(
'../save/1to7Attack/GDP_{}_{}_seed{}_clip{}_scale{}.txt'.format(
args.dataset, args.model, s, args.norm_bound, args.noise_scale),
testing_accuracy)
def federated_main(args):
# prepare logger
output_dir = "../logs"
if not os.path.exists(output_dir):
os.makedirs(output_dir)
session_time = strftime("%y%m%d-%H%M%S", localtime())
output_file = os.path.join(output_dir,
"{}-{}.log".format(session_time, args.dataset))
log_file = logging.FileHandler(output_file)
log_file.setFormatter(logging.Formatter('%(asctime)s: %(message)s'))
logger = logging.getLogger()
logger.setLevel(logging.INFO)
logger.addHandler(log_file)
# prepare client data output directory, related output is at bottom
if not os.path.exists("../clients_data"):
os.makedirs("../clients_data")
# prepare devices
if args.gpu:
torch.cuda.set_device("cuda:{}".format(args.gpu))
device = "cuda" if args.gpu else "cpu"
# prepare dataset
if args.dataset == "cifar":
dataset = CreateCIFAR10(args.num_users)
elif args.dataset == "mnist":
dataset = CreateMNIST(args.num_users)
elif args.dataset == "fmnist":
dataset = CreateMNIST(args.num_users, "../data/fmnist")
else:
raise ValueError("The dataset you input is not supported yet")
if args.iid:
user_groups = dataset.create_iid()
else:
if args.unequal:
if args.dataset == "cifar":
raise NotImplementedError(
"Cifar 10 unequal dataset is not implemented")
else:
user_groups = dataset.crate_noniid_unequal()
else:
user_groups = dataset.create_noniid()
# initialize the model
if args.model == "cnn":
if args.dataset == "mnist":
global_model = CNNMnist(args=args)
elif args.dataset == "fmnist":
global_model = CNNFashion_Mnist(args=args)
elif args.dataset == "cifar":
global_model = CNNCifar(args=args)
else:
raise ValueError("The dataset you input is not supported yet")
elif args.model == "mlp":
img_size = dataset.train_dataset.data[0].shape
len_in = 1
for x in img_size:
len_in *= x
global_model = MLP(dim_in=len_in,
dim_hidden=64,
dim_out=args.num_classes)
else:
raise ValueError("The model you input is not supported yet")
global_model.to(device)
global_model.train()
# prepare the parameters of global model
global_parameters = global_model.state_dict()
# Prepare the client container
local_outputs = [
ClientOutput(client_idx) for client_idx in range(args.num_users)
]
# malicious behavior, some nodes may perform model posion attack by changing the label of data
if args.malicious:
malicious_changer = MaliciousChange(args, dataset.train_dataset)
# the length of source_list and dest_list can influence the final accuracy
malicious_changer.data_poision(user_groups, [1, 5, 3], [3, 1, 5])
malicious_client_idxs = malicious_changer.malicious_clients
for idx in malicious_client_idxs:
local_outputs[idx].update(malicious=True)
# Prepare global early stopping
if args.global_early_stop:
global_early_stopping = EarlyStopping(
patience=args.global_early_rounds)
# Prepare the aggregation method
if args.aggre == "simple":
aggre_method = SimpleAggregation(int(args.rate * args.num_users),
int((args.num_users - 1) / 3), True,
args.no_consensus)
elif args.aggre == "time":
aggre_method = ThresholdAggregation(int(args.rate * args.num_users),
int((args.num_users - 1) / 3),
True, args.no_consensus)
elif args.aggre == "top":
aggre_method = TopAggregation(int(args.rate * args.num_users),
int((args.num_users - 1) / 3), False,
args.no_consensus)
else:
raise ValueError(
"The aggregation method you indicate has not been implemented")
# prepare other things
old_global_metric = 0.
server_total_profit = 0.
epoch_outputs = []
if args.malicious:
client_chooser = ChooseClient(args, malicious_client_idxs)
else:
client_chooser = ChooseClient(args)
# Training
for epoch in tqdm.tqdm(range(args.epochs)):
# Prepare some things
logger.info("Epoch {} Started".format(epoch))
# Choose clients participates in this epoch
if args.aggre == "simple":
# The number of received models will influence the performance
# In order to perform a fair comparison, FedAvg has no random screening
k = max(int(args.rate * args.num_users), 1)
else:
k = max(int(args.frac * args.num_users), 1)
chosen_clients = client_chooser.part_pick(list(range(args.num_users)),
k)
# Client part
for client_idx in chosen_clients:
# Client model training
client_updater = LocalUpdate(args, dataset.train_dataset,
user_groups[client_idx], logger)
client_model, client_loss = client_updater(
copy.deepcopy(global_model))
# Client model evaluation
evaluator = Evaluator(args, logger)
client_global_metric, client_global_loss = evaluator(
client_model, dataset.test_dataset)
# Update ClientOutput
local_outputs[client_idx].update(
parameter=copy.deepcopy(client_model.state_dict()),
number=len(user_groups[client_idx]),
epochs=epoch,
accuracy=client_global_metric,
local_loss=client_loss,
global_loss=client_global_loss)
epoch_outputs.append(local_outputs[client_idx])
# Update global parameters
epoch_outputs, global_parameters = aggre_method(epoch_outputs)
global_model.load_state_dict(global_parameters)
# Evaluate global model on global dataset
global_evaluator = Evaluator(args, logger)
global_metric, global_loss = global_evaluator(global_model,
dataset.test_dataset)
# Calculate the profit
epoch_outputs, publisher_profit, pay_count = epoch_incentive(
args, epoch_outputs, global_metric, old_global_metric)
server_total_profit += publisher_profit
# Epoch ends
logger.info("Epoch {} End".format(epoch))
# Print epoch information
logger.info("[client]Clients Information:")
for single_client in epoch_outputs:
client_dict = single_client.output()
client_idx = client_dict["idx"]
for client_key in client_dict:
if client_key != "idx":
logger.info("[client-{}]{}: {}".format(
client_idx, client_key, client_dict[client_key]))
logger.info("[server]Server information:")
logger.info("[server]loss: {}".format(global_loss))
logger.info("[server]accuracy: {}".format(global_metric))
logger.info("[server]epoch_profit: {}".format(
publisher_profit / pay_count if pay_count != 0 else 0))
logger.info("[server]total_profit: {}".format(server_total_profit))
# Global early stop mechanism
if args.global_early_stop and epoch > args.no_stoping_epochs:
global_early_stopping(global_loss, global_model)
if global_early_stopping.early_stop:
break
# Reset some values
# Only the best performance is meaningful to task participant
old_global_metric = max(global_metric, old_global_metric)
epoch_outputs = []
pickle.dump(
local_outputs,
open("../clients_data/{}-{}.pickle".format(session_time, args.dataset),
"wb"))
logger.info("Task finished!")
# Ensure correct weight is being updated
assert name == bl_name
# Calculate update
delta = weight - baseline
local_delta_update.append((name, delta))
local_weights.append(local_delta_update)
#local_weights.append(copy.deepcopy(w))
local_losses.append(copy.deepcopy(loss))
# update global weights
if epoch == 1: # initialize the global model by vanilla averaging for first epoch
global_weights = average_weights(w_s)
global_model.load_state_dict(global_weights)
else:
global_weights = fedavg(baseline_weights,
local_weights,
num_samples_list,
server_lr=server_lr)
updated_state_dict = {}
for name, weight in global_weights:
updated_state_dict[name] = weight
global_model.load_state_dict(updated_state_dict, strict=False)
loss_avg = sum(local_losses) / len(local_losses)
train_loss.append(loss_avg)
def main_test(args):
start_time = time.time()
now = datetime.datetime.now().strftime('%Y-%m-%d-%H%M%S')
# define paths
logger = SummaryWriter('../logs')
# easydict 사용하는 경우 주석처리
# args = args_parser()
# checkpoint 생성위치
args.save_path = os.path.join(args.save_path, args.exp_folder)
if not os.path.exists(args.save_path):
os.makedirs(args.save_path)
save_path_tmp = os.path.join(args.save_path, 'tmp_{}'.format(now))
if not os.path.exists(save_path_tmp):
os.makedirs(save_path_tmp)
SAVE_PATH = os.path.join(args.save_path, '{}_{}_T[{}]_C[{}]_iid[{}]_E[{}]_B[{}]'.
format(args.dataset, args.model, args.epochs, args.frac, args.iid,
args.local_ep, args.local_bs))
# 시드 고정
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
random.seed(args.seed)
np.random.seed(args.seed)
# torch.cuda.set_device(0)
device = torch.device("cuda:{}".format(args.gpu) if torch.cuda.is_available() else "cpu")
cpu_device = torch.device('cpu')
# log 파일 생성
log_path = os.path.join('../logs', args.exp_folder)
if not os.path.exists(log_path):
os.makedirs(log_path)
loggertxt = get_logger(
os.path.join(log_path, '{}_{}_{}_{}.log'.format(args.model, args.optimizer, args.norm, now)))
logging.info(args)
# csv
csv_save = '../csv/' + now
csv_path = os.path.join(csv_save, 'accuracy.csv')
csv_logger_keys = ['train_loss', 'accuracy']
csvlogger = CSVLogger(csv_path, csv_logger_keys)
# load dataset and user groups
train_dataset, test_dataset, client_loader_dict = get_dataset(args)
# cifar-100의 경우 자동 설정
if args.dataset == 'cifar100':
args.num_classes = 100
# BUILD MODEL
if args.model == 'cnn':
# Convolutional neural network
if args.dataset == 'mnist':
global_model = CNNMnist(args=args)
elif args.dataset == 'fmnist':
global_model = CNNFashion_Mnist(args=args)
elif args.dataset == 'cifar':
global_model = CNNCifar(args=args)
elif args.dataset == 'cifar100':
global_model = CNNCifar(args=args)
elif args.model == 'mlp':
# Multi-layer preceptron
img_size = train_dataset[0][0].shape
len_in = 1
for x in img_size:
len_in *= x
global_model = MLP(dim_in=len_in, dim_hidden=64,
dim_out=args.num_classes)
elif args.model == 'cnn_vc':
global_model = CNNCifar_fedVC(args=args)
elif args.model == 'cnn_vcbn':
global_model = CNNCifar_VCBN(args=args)
elif args.model == 'cnn_vcgn':
global_model = CNNCifar_VCGN(args=args)
elif args.model == 'resnet18_ws':
global_model = resnet18(num_classes=args.num_classes, weight_stand=1)
elif args.model == 'resnet18':
global_model = resnet18(num_classes=args.num_classes, weight_stand=0)
elif args.model == 'resnet32':
global_model = ResNet32_test(num_classes=args.num_classes)
elif args.model == 'resnet18_mabn':
global_model = resnet18_mabn(num_classes=args.num_classes)
elif args.model == 'vgg':
global_model = vgg11()
elif args.model == 'cnn_ws':
global_model = CNNCifar_WS(args=args)
else:
exit('Error: unrecognized model')
# Set the model to train and send it to device.
loggertxt.info(global_model)
# fedBN처럼 gn no communication 용
client_models = [copy.deepcopy(global_model) for idx in range(args.num_users)]
# copy weights
global_weights = global_model.state_dict()
global_model.to(device)
global_model.train()
# Training
train_loss, train_accuracy = [], []
val_acc_list, net_list = [], []
# how does help BN 확인용
client_loss = [[] for i in range(args.num_users)]
client_conv_grad = [[] for i in range(args.num_users)]
client_fc_grad = [[] for i in range(args.num_users)]
client_total_grad_norm = [[] for i in range(args.num_users)]
# 전체 loss 추적용 -how does help BN
# 재시작
if args.resume:
checkpoint = torch.load(SAVE_PATH)
global_model.load_state_dict(checkpoint['global_model'])
if args.hold_normalize:
for client_idx in range(args.num_users):
client_models[client_idx].load_state_dict(checkpoint['model_{}'.format(client_idx)])
else:
for client_idx in range(args.num_users):
client_models[client_idx].load_state_dict(checkpoint['global_model'])
resume_iter = int(checkpoint['a_iter']) + 1
print('Resume trainig form epoch {}'.format(resume_iter))
else:
resume_iter = 0
# learning rate scheduler
#scheduler = torch.optim.lr_scheduler.StepLR(optimizer=optimizer, gamma=0.1,step_size=500)
# start training
for epoch in tqdm(range(args.epochs)):
local_weights, local_losses = [], []
if args.verbose:
print(f'\n | Global Training Round : {epoch + 1} |\n')
global_model.train()
m = max(int(args.frac * args.num_users), 1)
idxs_users = np.random.choice(range(args.num_users), m, replace=False)
for idx in idxs_users:
"""
for key in global_model.state_dict().keys():
if args.hold_normalize:
if 'bn' not in key:
client_models[idx].state_dict()[key].data.copy_(global_model.state_dict()[key])
else:
client_models[idx].state_dict()[key].data.copy_(global_model.state_dict()[key])
"""
torch.cuda.empty_cache()
local_model = LocalUpdate(args=args, logger=logger, train_loader=client_loader_dict[idx], device=device)
w, loss, batch_loss, conv_grad, fc_grad, total_gard_norm = local_model.update_weights(
model=copy.deepcopy(global_model), global_round=epoch, idx_user=idx)
local_weights.append(copy.deepcopy(w))
# client의 1 epoch에서의 평균 loss값 ex)0.35(즉, batch loss들의 평균)
local_losses.append(copy.deepcopy(loss))
# 전체 round scheduler
# scheduler.step()
# loss graph용 -> client당 loss값 진행 저장 -> 모두 client별로 저장.
client_loss[idx].append(batch_loss)
client_conv_grad[idx].append(conv_grad)
client_fc_grad[idx].append(fc_grad)
client_total_grad_norm[idx].append(total_gard_norm)
# print(total_gard_norm)
# gn, bn 복사
# client_models[idx].load_state_dict(w)
del local_model
del w
# update global weights
global_weights = average_weights(local_weights, client_loader_dict, idxs_users)
# update global weights
# opt = OptRepo.name2cls('adam')(global_model.parameters(), lr=0.01, betas=(0.9, 0.99), eps=1e-3)
opt = OptRepo.name2cls('sgd')(global_model.parameters(), lr=10, momentum=0.9)
opt.zero_grad()
opt_state = opt.state_dict()
global_weights = aggregation(global_weights, global_model)
global_model.load_state_dict(global_weights)
opt = OptRepo.name2cls('sgd')(global_model.parameters(), lr=10, momentum=0.9)
# opt = OptRepo.name2cls('adam')(global_model.parameters(), lr=0.01, betas=(0.9, 0.99), eps=1e-3)
opt.load_state_dict(opt_state)
opt.step()
loss_avg = sum(local_losses) / len(local_losses)
train_loss.append(loss_avg)
global_model.eval()
# for c in range(args.num_users):
# local_model = LocalUpdate(args=args, dataset=train_dataset,
# idxs=user_groups[idx], logger=logger)
# acc, loss = local_model.inference(model=global_model)
# list_acc.append(acc)
# list_loss.append(loss)
# train_accuracy.append(sum(list_acc)/len(list_acc))
train_accuracy = test_inference(args, global_model, test_dataset, device=device)
val_acc_list.append(train_accuracy)
# print global training loss after every 'i' rounds
# if (epoch+1) % print_every == 0:
loggertxt.info(f' \nAvg Training Stats after {epoch + 1} global rounds:')
loggertxt.info(f'Training Loss : {loss_avg}')
loggertxt.info('Train Accuracy: {:.2f}% \n'.format(100 * train_accuracy))
csvlogger.write_row([loss_avg, 100 * train_accuracy])
if (epoch + 1) % 100 == 0:
tmp_save_path = os.path.join(save_path_tmp, 'tmp_{}.pt'.format(epoch+1))
torch.save(global_model.state_dict(),tmp_save_path)
# Test inference after completion of training
test_acc = test_inference(args, global_model, test_dataset, device=device)
print(' Saving checkpoints to {}...'.format(SAVE_PATH))
if args.hold_normalize:
client_dict = {}
for idx, model in enumerate(client_models):
client_dict['model_{}'.format(idx)] = model.state_dict()
torch.save(client_dict, SAVE_PATH)
else:
torch.save({'global_model': global_model.state_dict()}, SAVE_PATH)
loggertxt.info(f' \n Results after {args.epochs} global rounds of training:')
# loggertxt.info("|---- Avg Train Accuracy: {:.2f}%".format(100*train_accuracy[-1]))
loggertxt.info("|---- Test Accuracy: {:.2f}%".format(100 * test_acc))
# frac이 1이 아닐경우 잘 작동하지않음.
# batch_loss_list = np.array(client_loss).sum(axis=0) / args.num_users
# conv_grad_list = np.array(client_conv_grad).sum(axis=0) / args.num_users
# fc_grad_list = np.array(client_fc_grad).sum(axis=0) / args.num_users
# total_grad_list = np.array(client_total_grad_norm).sum(axis=0) /args.num_users
# client의 avg를 구하고 싶었으나 현재는 client 0만 확인
# client마다 batch가 다를 경우 bug 예상
return train_loss, val_acc_list, client_loss[0], client_conv_grad[0], client_fc_grad[0], client_total_grad_norm[0]
def main():
start_time = time.time()
# define paths
path_project = os.path.abspath('..')
logger = SummaryWriter('../logs')
args = args_parser()
exp_details(args)
if args.gpu:
torch.cuda.set_device(0)
device = 'cuda' if args.gpu else 'cpu'
# load dataset and user groups
train_dataset, test_dataset, user_groups = get_dataset(args)
args.num_users = len(user_groups)
# BUILD MODEL
if args.model == 'cnn':
# Convolutional neural netork
if args.dataset == 'mnist':
global_model = CNNMnist(args=args)
elif args.dataset == 'fmnist':
global_model = CNNFashion_Mnist(args=args)
elif args.dataset == 'cifar':
global_model = CNNCifar(args=args)
elif args.model == 'mlp':
# Multi-layer preceptron
img_size = train_dataset[0][0].shape
len_in = 1
for x in img_size:
len_in *= x
global_model = MLP(dim_in=len_in,
dim_hidden=64,
dim_out=args.num_classes)
else:
exit('Error: unrecognized model')
# Set the model to train and send it to device.
global_model.to(device)
global_model.train()
# copy weights
global_weights = global_model.state_dict()
# Training
train_loss, train_accuracy = [], []
val_acc_list, net_list = [], []
cv_loss, cv_acc = [], []
print_every = 2
val_loss_pre, counter = 0, 0
#Beolvassuk, hogy éppen mely résztvevők vesznek részt a tanításban (0 jelentése, hogy benne van, 1 az hogy nincs)
users = []
fp = open('users.txt', "r")
x = fp.readline().split(' ')
for i in x:
if i != '':
users.append(int(i))
fp.close()
#for epoch in tqdm(range(args.epochs)):
for epoch in range(args.epochs):
local_weights, local_losses = [], []
#print(f'\n | Global Training Round : {epoch+1} |\n')
global_model.train()
m = max(int(args.frac * args.num_users), 1)
idxs_users = np.random.choice(range(args.num_users), m, replace=False)
for idx in idxs_users:
local_model = LocalUpdate(args=args,
dataset=train_dataset,
idxs=user_groups[idx],
logger=logger)
w, loss = local_model.update_weights(
model=copy.deepcopy(global_model), global_round=epoch)
local_weights.append(copy.deepcopy(w))
local_losses.append(copy.deepcopy(loss))
global_weights = average_weights(local_weights)
# update global weights
global_model.load_state_dict(global_weights)
loss_avg = sum(local_losses) / len(local_losses)
train_loss.append(loss_avg)
# Calculate avg training accuracy over all users at every epoch
list_acc, list_loss = [], []
global_model.eval()
for c in range(args.num_users):
local_model = LocalUpdate(args=args,
dataset=train_dataset,
idxs=user_groups[idx],
logger=logger)
acc, loss = local_model.inference(model=global_model)
list_acc.append(acc)
list_loss.append(loss)
train_accuracy.append(sum(list_acc) / len(list_acc))
# print global training loss after every 'i' rounds
'''if (epoch+1) % print_every == 0:
print(f' \nAvg Training Stats after {epoch+1} global rounds:')
print(f'Training Loss : {np.mean(np.array(train_loss))}')
print('Train Accuracy: {:.2f}% \n'.format(100*train_accuracy[-1]))'''
# Test inference after completion of training
#Beolvassuk hogy mely résztvevőnek mely labeleket osztottuk ki.
ftrain = open('traindataset.txt')
testlabels = []
line = ftrain.readline()
while line != "":
sor = line.split(' ')
array = []
for i in sor:
array.append(int(i))
testlabels.append(array)
line = ftrain.readline()
ftrain.close()
print("USERS LABELS")
print(testlabels)
#Minden lehetséges koalícióra lefut a tesztelés
for j in range((2**args.num_users) - 1):
binary = numberToBinary(j, len(users))
test_acc, test_loss = test_inference(args, global_model, test_dataset,
testlabels, binary, len(binary))
#Teszt eredmények kiírása
print("RESZTVEVOK")
print(users)
print("TEST NUMBER")
print(j)
print("TEST BINARY")
print(binary)
print("TEST LABELS")
print(testlabels)
print("Test Accuracy")
print("{:.2f}%".format(100 * test_acc))
print()
# Saving the objects train_loss and train_accuracy:
'''file_name = '../save/objects/{}_{}_{}_C[{}]_iid[{}]_E[{}]_B[{}].pkl'.\
cwd = os.getcwd()
filepath = os.path.join(cwd,'save/objects/{}_{}_C[{}]_iid[{}]_B[{}]_N[{}]_MA[{}]_R[{}].pkl'.\
format(args.dataset, args.model, args.frac, args.iid,
args.local_bs, args.norm, args.ma, args.run))
# Logging setup
filename = os.path.join(cwd,'save/objects/{}_{}_{}_C[{}]_iid[{}]_E[{}]_B[{}]_N[{}]_MA[{}]_R[{}].log'.\
format(args.dataset, args.model, args.epochs, args.frac, args.iid,
args.local_ep, args.local_bs, args.norm, args.ma, args.run))
logging.basicConfig(filename=filename, filemode='a', level=logging.INFO)
if args.resume == 1:
state = load_model(filepath, args)
global_model.load_state_dict(state['state_dict'])
prev_epoch = state['epoch']
else:
prev_epoch = 0
# Set the model to train and send it to device.
global_model.to(device)
global_model.train()
print(global_model)
# copy weights
global_weights = global_model.state_dict()
# Training
train_loss, train_accuracy = [], []
test_loss, test_accuracy = [], []
def main():
start_time = time.time()
# define paths
path_project = os.path.abspath('..')
logger = SummaryWriter('../logs')
args = args_parser()
args = adatok.arguments(args)
exp_details(args)
if args.gpu:
torch.cuda.set_device(args.gpu)
device = 'cuda' if args.gpu else 'cpu'
# load dataset and user groups
train_dataset, test_dataset, user_groups = get_dataset(args)
if adatok.data.image_initialization == True:
adatok.data.image_initialization = False
return
# BUILD MODEL
if args.model == 'cnn':
# Convolutional neural netork
if args.dataset == 'mnist':
global_model = CNNMnist(args=args)
elif args.dataset == 'fmnist':
global_model = CNNFashion_Mnist(args=args)
elif args.dataset == 'cifar':
global_model = CNNCifar(args=args)
elif args.model == 'mlp':
# Multi-layer preceptron
img_size = train_dataset[0][0].shape
len_in = 1
for x in img_size:
len_in *= x
global_model = MLP(dim_in=len_in,
dim_hidden=64,
dim_out=args.num_classes)
else:
exit('Error: unrecognized model')
# Set the model to train and send it to device.
global_model.to(device)
global_model.train()
#print(global_model)
# copy weights
global_weights = global_model.state_dict()
# Training
train_loss, train_accuracy = [], []
val_acc_list, net_list = [], []
cv_loss, cv_acc = [], []
print_every = 2
val_loss_pre, counter = 0, 0
for epoch in tqdm(range(args.epochs)):
local_weights, local_losses = [], []
#print(f'\n | Global Training Round : {epoch+1} |\n')
global_model.train()
m = max(int(args.frac * args.num_users), 1)
idxs_users = np.random.choice(range(args.num_users), m, replace=False)
for idx in idxs_users:
local_model = LocalUpdate(args=args,
dataset=train_dataset,
idxs=user_groups[idx],
logger=logger)
w, loss = local_model.update_weights(
model=copy.deepcopy(global_model), global_round=epoch)
local_weights.append(copy.deepcopy(w))
local_losses.append(copy.deepcopy(loss))
# update global weights
global_weights = average_weights(local_weights)
# update global weights
global_model.load_state_dict(global_weights)
loss_avg = sum(local_losses) / len(local_losses)
train_loss.append(loss_avg)
# Calculate avg training accuracy over all users at every epoch
list_acc, list_loss = [], []
global_model.eval()
for c in range(args.num_users):
local_model = LocalUpdate(args=args,
dataset=train_dataset,
idxs=user_groups[idx],
logger=logger)
acc, loss = local_model.inference(model=global_model)
list_acc.append(acc)
list_loss.append(loss)
train_accuracy.append(sum(list_acc) / len(list_acc))
# print global training loss after every 'i' rounds
'''if (epoch+1) % print_every == 0:
print(f' \nAvg Training Stats after {epoch+1} global rounds:')
print(f'Training Loss : {np.mean(np.array(train_loss))}')
print('Train Accuracy: {:.2f}% \n'.format(100*train_accuracy[-1]))'''
# Test inference after completion of training
for i in adatok.data.test_groups_in_binary:
adatok.data.actual_test_group_in_binary = i
test_acc, test_loss = test_inference(args, global_model,
test_dataset)
print("Resoults")
print(epoch)
print(adatok.data.actual_train_group_in_binary)
print(adatok.data.actual_test_group_in_binary)
print(test_acc)
print(test_loss)
'''
def poisoned_NoDefense(nb_attackers, seed=1):
# define paths
path_project = os.path.abspath('..')
logger = SummaryWriter('../logs')
args = args_parser()
exp_details(args)
# set seed
torch.manual_seed(seed)
np.random.seed(seed)
# device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# load dataset and user groups
train_dataset, test_dataset, user_groups = get_dataset(args)
# BUILD MODEL
if args.model == 'cnn':
# Convolutional neural netork
if args.dataset == 'mnist':
global_model = CNNMnist(args=args)
elif args.dataset == 'fmnist':
global_model = CNNFashion_Mnist(args=args)
elif args.dataset == 'cifar':
global_model = CNNCifar(args=args)
elif args.model == 'mlp':
# Multi-layer preceptron
img_size = train_dataset[0][0].shape
len_in = 1
for x in img_size:
len_in *= x
global_model = MLP(dim_in=len_in, dim_hidden=64,
dim_out=args.num_classes)
else:
exit('Error: unrecognized model')
# Set the model to train and send it to device.
global_model.to(device)
global_model.train()
print(global_model)
# copy weights
global_weights = global_model.state_dict()
# backdoor model
dummy_model = copy.deepcopy(global_model)
dummy_model.load_state_dict(torch.load('../save/all_5_model.pth'))
dummy_norm = 0
for x in dummy_model.state_dict().values():
dummy_norm += x.norm(2).item() ** 2
dummy_norm = dummy_norm ** (1. / 2)
# testing accuracy for global model
testing_accuracy = [0.1]
for epoch in tqdm(range(args.epochs)):
local_del_w = []
print(f'\n | Global Training Round : {epoch+1} |\n')
global_model.train()
m = max(int(args.frac * args.num_users), 1)
idxs_users = np.random.choice(range(args.num_users), m, replace=False)
# Adversary updates
for idx in idxs_users[0:nb_attackers]:
print("evil")
local_model = LocalUpdate(args=args, dataset=train_dataset, idxs=user_groups[idx], logger=logger)
#del_w, _ = local_model.poisoned_SGA(model=copy.deepcopy(global_model), change=1)
w = copy.deepcopy(dummy_model)
# compute change in parameters and norm
zeta = 0
for del_w, w_old in zip(w.parameters(), global_model.parameters()):
del_w.data -= copy.deepcopy(w_old.data)
del_w.data *= m / nb_attackers
del_w.data += copy.deepcopy(w_old.data)
zeta += del_w.norm(2).item() ** 2
zeta = zeta ** (1. / 2)
del_w = copy.deepcopy(w.state_dict())
local_del_w.append(copy.deepcopy(del_w))
# Non-adversarial updates
for idx in idxs_users[nb_attackers:]:
print("good")
local_model = LocalUpdate(args=args, dataset=train_dataset, idxs=user_groups[idx], logger=logger)
del_w, _ = local_model.update_weights(model=copy.deepcopy(global_model), change=1)
local_del_w.append(copy.deepcopy(del_w))
# average local updates
average_del_w = average_weights(local_del_w)
# Update global model: w_{t+1} = w_{t} + average_del_w
for param, param_del_w in zip(global_weights.values(), average_del_w.values()):
param += param_del_w
global_model.load_state_dict(global_weights)
# test accuracy
test_acc, test_loss = test_inference(args, global_model, test_dataset)
testing_accuracy.append(test_acc)
print("Test accuracy")
print(testing_accuracy)
# save test accuracy
np.savetxt('../save/RandomAttack/NoDefense_iid_{}_{}_attackers{}_seed{}.txt'.
format(args.dataset, args.model, nb_attackers, s), testing_accuracy)