Cain：提前写一下AI部分的WP

明后天统计队内师傅的WP，更新全部赛题

WriteUP

羊城杯AI

AI-NLP_Model_Attack

关于bert的llm对抗，使用清华的openattack就行，使用了基于贪婪算法进行word词替换的PWWS攻击

import OpenAttack as oa
import numpy as np
import torch
import datasets
import transformers
from transformers import DistilBertTokenizer, DistilBertForSequenceClassification


import csv
from collections import defaultdict
with open('original_text.csv', mode='r', encoding='utf-8') as file:
    # 创建CSV阅读器
    reader = csv.DictReader(file)
    
    # 创建一个字典来存储转换后的数据
    data_dict = {
            "x": [
                ],
            "y": [
            ]
    }
    
    # 遍历CSV文件中的每一行
    for row in reader:
        # 将'text'列的值添加到data_dict的'x'键下
        data_dict['x'].append(row['text'])
        # 将'original_label'列的值添加到data_dict的'y'键下
        data_dict['y'].append(int(row['original_label']))  # 假设original_label是整数

tokenizer = DistilBertTokenizer.from_pretrained('Sentiment_classification_model')
model = DistilBertForSequenceClassification.from_pretrained('Sentiment_classification_model').to('cuda')
victim = oa.classifiers.TransformersClassifier(model, tokenizer, model.distilbert.embeddings.word_embeddings, 'cuda')

attacker = oa.attackers.PWWSAttacker()
dataset = datasets.Dataset.from_dict(data_dict)
attack_eval = oa.AttackEval(attacker, victim, metrics = [oa.metric.EditDistance(), oa.metric.ModificationRate()])
attack_eval.eval(dataset, visualize=True)

AI-Targeted_Image_adv_attacks

题目要求进行有目标的对抗样本攻击，并且检验方式为SIMM

因此对损失函数使用SIMM进行约束，在SIMM条件下的进行约束优化，采用了梯度优化FGSM

考虑到模型敏感度使用softmax后进行交叉熵的效果并不好，观察模型输出后，直接预设期望的target计算MSE均方误差

from torch import nn
import torch.nn.functional as func
import numpy as np
import torch
from PIL import Image
import base64
import random
from torchvision import models, transforms
import pytorch_ssim
from torch.autograd import Variable

label = {
    0: 'cat',
    1: 'dog',
    2: 'fox'
}
 
target = [torch.tensor([[-10, 10, -10]]).to('cuda').float(),
          torch.tensor([[-10, -10, 10]]).to('cuda').float(),
          torch.tensor([[10, -10, -10]]).to('cuda').float()]

model = models.densenet121(weights=None)
num_ftrs = model.classifier.in_features
model.classifier = nn.Sequential(
    nn.Linear(num_ftrs, 500),
    nn.Linear(500, 3) 
)

model.load_state_dict(torch.load('ckpt_densenet121_catdogfox_classify.pth', map_location='cuda'))
model = model.to('cuda')
model.eval()
loss_ssim = pytorch_ssim.SSIM()
loss_l2 = nn.MSELoss()
delta = 0.01
cnt = 0

for idx in range(3):

    for num in range(50):

        print(f'--------{label[idx]}_{num:03d} Start Adv Attack!---------')

        img = transforms.ToTensor()(Image.open(f'.\adv_image\{label[idx]}\{label[idx]}_{num:03d}.jpg')).to('cuda').unsqueeze(0)
        src_img = transforms.ToTensor()(Image.open(f'.\adv_image\{label[idx]}\{label[idx]}_{num:03d}.jpg')).to('cuda').unsqueeze(0)
        src = src_img.view(-1).detach().cpu().numpy()

        img.requires_grad = True

        for i in range(100000):
            pred = model(img)

            print(f"Epoch {i}: target pred: {pred[0].argmax().item()}",end='  ')

            adv = transforms.ToPILImage()(img[0].detach().cpu())
            adv.save(f'.\update\{label[idx]}\{label[idx]}_{num:03d}.jpg')
            adv_img = transforms.ToTensor()(Image.open(f'.\update\{label[idx]}\{label[idx]}_{num:03d}.jpg')).to('cuda').unsqueeze(0)
            
            if pytorch_ssim.ssim(adv_img, src_img) < 0.95:
                print(f'-------- {label[idx]}_{num:03d}  SSIM < 0.95---------')
                cnt += 1
                break

            if model(adv_img)[0].argmax().item() == (idx+1)%3 and pytorch_ssim.ssim(adv_img, src_img) > 0.95:
                print('--------Find Adv Image!---------')
                print(pytorch_ssim.ssim(img, src_img), pred[0].argmax().item())
                break

            ssiml = loss_ssim(img, src_img)
            l2l = loss_l2(pred, target[idx])

            loss = l2l + ssiml
            loss.backward()

            img.requires_grad = False

            img = img - img.grad * delta
            img = img.clamp(0, 1)

            img.requires_grad = True
            print(f"SIMM loss: {ssiml.item()}  MSE norm: {l2l.item()}")

print(f'-------- Total:{cnt} SSIM < 0.95---------')