image_train.py

import utils.csv_record as csv_record
import torch
import torch.nn as nn
import torch.nn.functional as F
import time
import main
import test
import copy
import config


def ImageTrain(helper, start_epoch, local_model, target_model, is_poison,agent_name_keys):

    epochs_submit_update_dict = dict()
    num_samples_dict = dict()
    current_number_of_adversaries=0
    #当前带后门C端数量
    for temp_name in agent_name_keys:
        if temp_name in helper.params['total_list']:
            current_number_of_adversaries+=1

    for model_id in range(helper.params['no_models']):
        epochs_local_update_list = []
        last_local_model = dict()
        client_grad = [] # only works for aggr_epoch_interval=1


        for name, data in target_model.state_dict().items():
            last_local_model[name] = target_model.state_dict()[name].clone()

        agent_name_key = agent_name_keys[model_id]
        ## Synchronize LR and models
        model = local_model
        model.copy_params(target_model.state_dict())
        optimizer = torch.optim.SGD(model.parameters(), lr=helper.params['lr'],
                                    momentum=helper.params['momentum'],
                                    weight_decay=helper.params['decay'])
        model.train()
        adversarial_index= -1
        localmodel_poison_epochs = helper.params['poison_epochs']
        #查询当前是否是注入后门的轮次，打印当前的带后门C端的索引
        if is_poison and agent_name_key in helper.params['total_list']:
            for temp_index in range(0, len(helper.params['total_list'])):
                if int(agent_name_key) == helper.params['total_list'][temp_index]:
                    adversarial_index= temp_index
                    localmodel_poison_epochs = helper.params[str(temp_index) + '_poison_epochs']
                    main.logger.info(
                        f'当前第{adversarial_index}位带后门C端：{agent_name_key} ')
                    break
            if len(helper.params['total_list']) == 1:
                adversarial_index = -1  # the global pattern

        for epoch in range(start_epoch, start_epoch + helper.params['aggr_epoch_interval']):

            target_params_variables = dict()
            for name, param in target_model.named_parameters():
                target_params_variables[name] = last_local_model[name].clone().detach().requires_grad_(False)
            #单发攻击，只在特定轮次攻击
            if is_poison and agent_name_key in helper.params['total_list'] and (epoch in localmodel_poison_epochs):
                main.logger.info('当前轮次将注入后门')

                poison_lr = helper.params['poison_lr']
                internal_epoch_num = helper.params['internal_poison_epochs']
                step_lr = helper.params['poison_step_lr']

                poison_optimizer = torch.optim.SGD(model.parameters(), lr=poison_lr,
                                                   momentum=helper.params['momentum'],
                                                   weight_decay=helper.params['decay'])
                scheduler = torch.optim.lr_scheduler.MultiStepLR(poison_optimizer,milestones=[3,5], gamma=0.1)
                temp_local_epoch = (epoch - 1) *internal_epoch_num
                for internal_epoch in range(1, internal_epoch_num + 1):
                    temp_local_epoch += 1
                    _, data_iterator = helper.train_data[agent_name_key]
                    poison_data_count = 0
                    total_loss = 0.
                    correct = 0
                    dataset_size = 0
                    dis2global_list=[]
                    for batch_id, batch in enumerate(data_iterator):
                        data, targets, poison_num = helper.get_poison_batch(batch, adversarial_index=adversarial_index,evaluation=False)
                        poison_optimizer.zero_grad()
                        dataset_size += len(data)
                        poison_data_count += poison_num

                        output = model(data)
                        class_loss = nn.functional.cross_entropy(output, targets)

                        distance_loss = helper.model_dist_norm_var(model, target_params_variables)
                        # Lmodel = αLclass + (1 − α)Lano; alpha_loss =1 fixed
                        loss = helper.params['alpha_loss'] * class_loss + \
                               (1 - helper.params['alpha_loss']) * distance_loss
                        loss.backward()

                        # get gradients
                        if helper.params['aggregation_methods']==config.AGGR_FOOLSGOLD:
                            for i, (name, params) in enumerate(model.named_parameters()):
                                if params.requires_grad:
                                    if internal_epoch == 1 and batch_id == 0:
                                        client_grad.append(params.grad.clone())
                                    else:
                                        client_grad[i] += params.grad.clone()

                        poison_optimizer.step()
                        total_loss += loss.data
                        pred = output.data.max(1)[1]  # get the index of the max log-probability
                        correct += pred.eq(targets.data.view_as(pred)).cpu().sum().item()

                        if helper.params["batch_track_distance"]:  #false
                            # we can calculate distance to this model now.
                            temp_data_len = len(data_iterator)
                            distance_to_global_model = helper.model_dist_norm(model, target_params_variables)
                            dis2global_list.append(distance_to_global_model)
                            model.track_distance_batch_vis(vis=main.vis, epoch=temp_local_epoch,
                                                           data_len=temp_data_len,
                                                            batch=batch_id,distance_to_global_model= distance_to_global_model,
                                                           eid=helper.params['environment_name'],
                                                           name=str(agent_name_key),is_poisoned=True)

                    if step_lr:
                        scheduler.step()
                        main.logger.info(f'当前学习率: {scheduler.get_last_lr()}')

                    acc = 100.0 * (float(correct) / float(dataset_size))
                    total_l = total_loss / dataset_size
                    main.logger.info(
                        '模型名称： {} ,  当前轮次： {:3d}, C端名称： {}, 目前处于多少轮 {:3d},  平均损失: {:.4f}, '
                        '准确率: {}/{} ({:.4f}%), 参与训练的数据总量: {}'.format(model.name, epoch, agent_name_key,
                                                                                      internal_epoch,
                                                                                      total_l, correct, dataset_size,
                                                                                     acc, poison_data_count))
                    csv_record.train_result.append(
                        [agent_name_key, temp_local_epoch,
                         epoch, internal_epoch, total_l.item(), acc, correct, dataset_size])
                    if helper.params['vis_train']:
                        model.train_vis(main.vis, temp_local_epoch,
                                        acc, loss=total_l, eid=helper.params['environment_name'], is_poisoned=True,
                                        name=str(agent_name_key) )
                    num_samples_dict[agent_name_key] = dataset_size
                    if helper.params["batch_track_distance"]:
                        main.logger.info(
                            f'MODEL {model_id}. P-norm is {helper.model_global_norm(model):.4f}. '
                            f'Distance to the global model: {dis2global_list}. ')

                # internal epoch finish
                main.logger.info(f'Global model norm: {helper.model_global_norm(target_model)}.')
                main.logger.info(f'Norm before scaling: {helper.model_global_norm(model)}. '
                                 f'Distance: {helper.model_dist_norm(model, target_params_variables)}')

                if not helper.params['baseline']:
                    main.logger.info(f'当前本地模型的主任务以及触发成功率：')
                    epoch_loss, epoch_acc, epoch_corret, epoch_total = test.Mytest(helper=helper, epoch=epoch,
                                                                                   model=model, is_poison=False,
                                                                                   visualize=False,
                                                                                   agent_name_key=agent_name_key)
                    csv_record.test_result.append(
                        [agent_name_key, epoch, epoch_loss, epoch_acc, epoch_corret, epoch_total])

                    epoch_loss, epoch_acc, epoch_corret, epoch_total = test.Mytest_poison(helper=helper,
                                                                                          epoch=epoch,
                                                                                          model=model,
                                                                                          is_poison=True,
                                                                                          visualize=False,
                                                                                          agent_name_key=agent_name_key)
                    csv_record.posiontest_result.append(
                        [agent_name_key, epoch, epoch_loss, epoch_acc, epoch_corret, epoch_total])

                    clip_rate = helper.params['scale_weights_poison']
                    main.logger.info(f"Scaling by  {clip_rate}")
                    for key, value in model.state_dict().items():
                        target_value  = last_local_model[key]
                        new_value = target_value + (value - target_value) * clip_rate
                        model.state_dict()[key].copy_(new_value)
                    distance = helper.model_dist_norm(model, target_params_variables)
                    main.logger.info(
                        f'注入后门后的距离参数: '
                        f'{helper.model_global_norm(model)}, distance: {distance}')
                    csv_record.scale_temp_one_row.append(epoch)
                    csv_record.scale_temp_one_row.append(round(distance, 4))
                    if helper.params["batch_track_distance"]: #false
                        temp_data_len = len(helper.train_data[agent_name_key][1])
                        model.track_distance_batch_vis(vis=main.vis, epoch=temp_local_epoch,
                                                       data_len=temp_data_len,
                                                       batch=temp_data_len-1,
                                                       distance_to_global_model=distance,
                                                       eid=helper.params['environment_name'],
                                                       name=str(agent_name_key), is_poisoned=True)
                if helper.params['diff_privacy']:
                    model_norm = helper.model_dist_norm(model, target_params_variables)  # 求模型范数

                    # 该模型范数若超过helper中的s_norm大小，则需要进行缩放
                    if model_norm > helper.params['s_norm']:
                        norm_scale = helper.params['s_norm'] / (model_norm)  # 缩放比例
                        for name, layer in model.named_parameters():
                            #### don't scale tied weights:
                            if helper.params.get('tied', False) and name == 'decoder.weight' or '__' in name:
                                continue

                            # 同269行
                            clipped_difference = norm_scale * (
                                    layer.data - target_model.state_dict()[name])

                            # 同273行
                            layer.data.copy_(target_model.state_dict()[name] + clipped_difference)
                distance = helper.model_dist_norm(model, target_params_variables)
                main.logger.info(f"Total norm for {current_number_of_adversaries} "
                                 f"adversaries is: {helper.model_global_norm(model)}. distance: {distance}")

            else:
                temp_local_epoch = (epoch - 1) * helper.params['internal_epochs']
                for internal_epoch in range(1, helper.params['internal_epochs'] + 1):
                    temp_local_epoch += 1

                    _, data_iterator = helper.train_data[agent_name_key]
                    total_loss = 0.
                    correct = 0
                    dataset_size = 0
                    dis2global_list = []
                    for batch_id, batch in enumerate(data_iterator):

                        optimizer.zero_grad()
                        data, targets = helper.get_batch(data_iterator, batch,evaluation=False)

                        dataset_size += len(data)
                        output = model(data)
                        loss = nn.functional.cross_entropy(output, targets)
                        loss.backward()

                        # get gradients
                        if helper.params['aggregation_methods'] == config.AGGR_FOOLSGOLD:
                            for i, (name, params) in enumerate(model.named_parameters()):
                                if params.requires_grad:
                                    if internal_epoch == 1 and batch_id == 0:
                                        client_grad.append(params.grad.clone())
                                    else:
                                        client_grad[i] += params.grad.clone()
                        
                        if helper.params['diff_privacy']:
                            optimizer.step()  # 经典3
                            model_norm = helper.model_dist_norm(model, target_params_variables)  # 求模型范数
                            main.logger.info(f'Test_model_dist_norm: {model_norm}')
                            # 如果范数过大，则进行“裁剪”（这个一块代码块复用前方）
                            if model_norm > helper.params['s_norm']:
                                norm_scale = helper.params['s_norm'] / (model_norm)
                                for name, layer in model.named_parameters():
                                    #### don't scale tied weights:
                                    if helper.params.get('tied',
                                                         False) and name == 'decoder.weight' or '__' in name:
                                        continue
                                    clipped_difference = norm_scale * (
                                            layer.data - target_model.state_dict()[name])
                                    layer.data.copy_(
                                        target_model.state_dict()[name] + clipped_difference)
                        else:
                            optimizer.step()
                        total_loss += loss.data
                        pred = output.data.max(1)[1]  # get the index of the max log-probability
                        correct += pred.eq(targets.data.view_as(pred)).cpu().sum().item()

                        if helper.params["vis_train_batch_loss"]:
                            cur_loss = loss.data
                            temp_data_len = len(data_iterator)
                            model.train_batch_vis(vis=main.vis,
                                                  epoch=temp_local_epoch,
                                                  data_len=temp_data_len,
                                                  batch=batch_id,
                                                  loss=cur_loss,
                                                  eid=helper.params['environment_name'],
                                                  name=str(agent_name_key) , win='train_batch_loss', is_poisoned=False)
                        if helper.params["batch_track_distance"]:
                            # we can calculate distance to this model now
                            temp_data_len = len(data_iterator)
                            distance_to_global_model = helper.model_dist_norm(model, target_params_variables)
                            dis2global_list.append(distance_to_global_model)
                            model.track_distance_batch_vis(vis=main.vis, epoch=temp_local_epoch,
                                                           data_len=temp_data_len,
                                                            batch=batch_id,distance_to_global_model= distance_to_global_model,
                                                           eid=helper.params['environment_name'],
                                                           name=str(agent_name_key),is_poisoned=False)

                    acc = 100.0 * (float(correct) / float(dataset_size))
                    total_l = total_loss / dataset_size
                    main.logger.info('___Train 模型名称： {} ,  当前轮次： {:3d}, C端名称： {}, 目前处于多少轮 {:3d},  平均损失: {:.4f}, '
                        '准确率: {}/{} ({:.4f}%)'.format(model.name, epoch, agent_name_key, internal_epoch,
                                                           total_l, correct, dataset_size,acc)
                    )

                    csv_record.train_result.append([agent_name_key, temp_local_epoch,
                                                    epoch, internal_epoch, total_l.item(), acc, correct, dataset_size])

                    if helper.params['vis_train']:
                        model.train_vis(main.vis, temp_local_epoch,
                                        acc, loss=total_l, eid=helper.params['environment_name'], is_poisoned=False,
                                        name=str(agent_name_key))
                    num_samples_dict[agent_name_key] = dataset_size

                    if helper.params["batch_track_distance"]:
                        main.logger.info(
                            f'MODEL {model_id}. P-norm is {helper.model_global_norm(model):.4f}. '
                            f'Distance to the global model: {dis2global_list}. ')

                # test local model after internal epoch finishing
                epoch_loss, epoch_acc, epoch_corret, epoch_total = test.Mytest(helper=helper, epoch=epoch,
                                                                               model=model, is_poison=False, visualize=True,
                                                                               agent_name_key=agent_name_key)
                csv_record.test_result.append([agent_name_key, epoch, epoch_loss, epoch_acc, epoch_corret, epoch_total])

            if is_poison:
                if agent_name_key in helper.params['total_list'] and (epoch in localmodel_poison_epochs):
                    epoch_loss, epoch_acc, epoch_corret, epoch_total = test.Mytest_poison(helper=helper,
                                                                                          epoch=epoch,
                                                                                          model=model,
                                                                                          is_poison=True,
                                                                                          visualize=True,
                                                                                          agent_name_key=agent_name_key)
                    csv_record.posiontest_result.append(
                        [agent_name_key, epoch, epoch_loss, epoch_acc, epoch_corret, epoch_total])

                #  test on local triggers
                if agent_name_key in helper.params['adversary_list_1']:
                    if helper.params['vis_trigger_split_test']:
                        model.trigger_agent_test_vis(vis=main.vis, epoch=epoch, acc=epoch_acc, loss=None,
                                                     eid=helper.params['environment_name'],
                                                     name=str(agent_name_key)  + "_combine")

                    epoch_loss, epoch_acc, epoch_corret, epoch_total = \
                        test.Mytest_poison_agent_trigger(helper=helper, model=model, agent_name_key=agent_name_key)
                    main.logger.info(f'误触率： 当前轮次:{epoch} 本轮的参与方 : {agent_name_key}   平均损失 : {epoch_loss}  正确率 : {epoch_acc}.')
                    csv_record.poisontriggertest_result.append(
                        [agent_name_key, str(agent_name_key) + "_trigger", "", epoch, epoch_loss,
                         epoch_acc, epoch_corret, epoch_total])
                    if helper.params['vis_trigger_split_test']:
                        model.trigger_agent_test_vis(vis=main.vis, epoch=epoch, acc=epoch_acc, loss=None,
                                                     eid=helper.params['environment_name'],
                                                     name=str(agent_name_key) + "_trigger")

            # update the model weight
            local_model_update_dict = dict()
            for name, data in model.state_dict().items():
                local_model_update_dict[name] = torch.zeros_like(data)
                local_model_update_dict[name] = (data - last_local_model[name])
                last_local_model[name] = copy.deepcopy(data)

            if helper.params['aggregation_methods'] == config.AGGR_FOOLSGOLD:
                epochs_local_update_list.append(client_grad)
            else:
                epochs_local_update_list.append(local_model_update_dict)

        epochs_submit_update_dict[agent_name_key] = epochs_local_update_list

    return epochs_submit_update_dict, num_samples_dict