diff options
| -rw-r--r-- | attack/excute.py | 93 |
1 files changed, 93 insertions, 0 deletions
diff --git a/attack/excute.py b/attack/excute.py new file mode 100644 index 0000000..9fb07e7 --- /dev/null +++ b/attack/excute.py @@ -0,0 +1,93 @@ +""" +Date: 2022-03-08 +Author: [email protected] +Desc: execute the adversarial attack +""" + +from tqdm import tqdm +import torch +from sklearn.metrics import accuracy_score +from sklearn.metrics import recall_score +from sklearn.metrics import f1_score +from sklearn.metrics import confusion_matrix +from TargetModel.FSNet.FSNet import FSNet +from attack.adversarialDataset import AdversarialC2Data +from torch.utils.data import DataLoader + +def executeAttack(dataloader, model, device, file): + y_true = [] + y_pred = [] + for batch_x, batch_y in tqdm(dataloader): + batch_x = batch_x.to(device) + batch_y = batch_y.to(device) + y_true += batch_y.detach().cpu().numpy().reshape(-1).tolist() + output = model(batch_x) + y_pred += torch.argmax(output, dim=1).detach().cpu().numpy().tolist() + acc = accuracy_score(y_true, y_pred) + recall = recall_score(y_true, y_pred, average='macro') + f1 = f1_score(y_true, y_pred, average='macro') + con_matrix = confusion_matrix(y_true, y_pred) + print("acc: {:.2%}, recall: {:.2%}, f1: {:.2%}".format(acc, recall, f1)) + print(con_matrix) + # print("acc: {:.2%}, recall: {:.2%}, f1: {:.2%}".format(acc, recall, f1), file=file) + # print(con_matrix, file=file) + return acc, recall, f1, con_matrix + + + + +def attackFsnet(theta, gamma, file, advdataFile, modelfile, device): + state = torch.load(modelfile) + fsnet = FSNet(state['param']) + fsnet.load_state_dict(state['model_dict']) + fsnet.to(device) + + # adversarial data + batch_size = 128 + adversarialC2Data = AdversarialC2Data(advdataFile, keep_target=True) + adversarialDataloader = DataLoader(adversarialC2Data, batch_size, shuffle=True, drop_last=False) + print("Attack FSNet") + print(fsnet) + + fsnet.eval() + y_true = [] + y_pred = [] + for batch_x, batch_y in tqdm(adversarialDataloader): + batch_x = batch_x.to(device) + batch_y = batch_y.to(device) + # batch_x.shape: (batch_size, sequenceLen) + # batch_y.shape: (batch_size, 1) + # reconstruction + z_e = fsnet.encode(batch_x) + # z_e.shape: (batch_size, num_layers * num_direction, hidden_size) + z_d, D = fsnet.decode(z_e) + # z_d.shape: (batch_size, num_layers * num_direction, hidden_size) + # D.shape: (batch_size, sequence_len, hidden_size * num_direction) + z_reconstruction = fsnet.reconstruction(D) + # z_reconstruction.shape=(batch_size, sequence_len, vocab_size) + + # classification layer + z_dense = fsnet.dense(z_e, z_d) + # z_dense.shape=(batch_size, num_class) + + # compute reconstruction loss + # z_reconstruction.shape = (-1, vocab_size) + # batch_x.shape=(-1) + + # compute classification loss + # batch_y.shape=(batch, ) + + # compute FPR + y_true += batch_y.detach().cpu().data.numpy().reshape(-1).tolist() + y_pred += torch.argmax(z_dense, dim=1).detach().cpu().data.numpy().tolist() + acc = accuracy_score(y_true, y_pred) + recall = recall_score(y_true, y_pred, average='macro') + f1 = f1_score(y_true, y_pred, average='macro') + con_matrix = confusion_matrix(y_true, y_pred) + print("acc: {:.2%}, recall: {:.2%}, f1: {:.2%}".format(acc, recall, f1)) + print(con_matrix) + print("acc: {:.2%}, recall: {:.2%}, f1: {:.2%}".format(acc, recall, f1), file=file) + print(con_matrix, file=file) + return con_matrix + + |