SMART/test/test_cw_attack.py

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision
import torchvision.transforms as transforms
from utils.carlini_wagner_l2 import carlini_wagner_l2
import matplotlib.pyplot as plt
from tqdm import tqdm
import os

# 设置随机种子以便结果可重现
def set_seed(seed):
    torch.manual_seed(seed)
    torch.cuda.manual_seed(seed)
    torch.cuda.manual_seed_all(seed)
    np.random.seed(seed)
    torch.backends.cudnn.deterministic = True
    torch.backends.cudnn.benchmark = False

set_seed(42)

# 设置设备
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using device: {device}")

# 加载CIFAR-10数据集
def load_cifar10():
    transform_test = transforms.Compose([
        transforms.ToTensor(),
    ])
    testset = torchvision.datasets.CIFAR10(root='./data', train=False,
                                          download=True, transform=transform_test)
    testloader = torch.utils.data.DataLoader(testset, batch_size=1,
                                          shuffle=False, num_workers=2)
    return testloader

# 加载ResNet50模型
def load_model():
    model = torchvision.models.resnet50(pretrained=True)
    
    # 修改最后的全连接层以适应CIFAR-10的10个类别
    num_ftrs = model.fc.in_features
    model.fc = nn.Linear(num_ftrs, 10)
    
    # 尝试加载预训练的CIFAR-10权重
    try:
        weight_path = "/share/pretrained_weights/cifar10_resnet50_cross_entropy.model"
        if os.path.exists(weight_path):
            state_dict = torch.load(weight_path)
            model.load_state_dict(state_dict)
            print(f"Loaded pretrained weights from {weight_path}")
        else:
            print(f"Warning: Pretrained weights not found at {weight_path}")
            print("Using ImageNet pretrained weights with modified final layer")
    except Exception as e:
        print(f"Error loading pretrained weights: {e}")
    
    model = model.to(device)
    model.eval()
    return model

# 对单个样本执行CW攻击
def test_cw_attack_on_sample(model, inputs, labels):
    model.eval()
    inputs = inputs.to(device)
    labels = labels.to(device)
    
    # 获取原始预测
    with torch.no_grad():
        outputs = model(inputs)
        _, predicted = torch.max(outputs, 1)
    
    print(f"Original prediction: {predicted.item()}, True label: {labels.item()}")
    print(f"Model output shape: {outputs.shape}")
    print(f"Confidence: {F.softmax(outputs, dim=1).max().item():.4f}")
    
    # 将模型包装为一个函数
    def model_fn(x):
        return model(x)
    
    # 执行CW攻击
    print("Running CW attack...")
    try:
        adv_inputs = carlini_wagner_l2(
            model_fn=model_fn,
            x=inputs,
            n_classes=10,  # CIFAR-10有10个类别
            y=predicted,   # 使用模型预测的类别
            targeted=False,  # 使用untargeted攻击
            lr=0.005,       # 学习率
            binary_search_steps=5,  # 二分搜索步骤
            max_iterations=500,     # 最大迭代次数
            confidence=0,      # 默认置信度
            clip_min=0.0,      # 图像范围限制
            clip_max=1.0,      # 图像范围限制
            early_stop=True    # 启用提前停止
        )
        
        # 获取对抗样本的预测
        with torch.no_grad():
            adv_outputs = model(adv_inputs)
            _, adv_predicted = torch.max(adv_outputs, 1)
        
        # 计算扰动大小(L2范数)
        perturbation = adv_inputs - inputs
        l2_norm = torch.norm(perturbation.view(perturbation.shape[0], -1), p=2, dim=1)
        
        print(f"Adversarial prediction: {adv_predicted.item()}")
        print(f"Attack success: {adv_predicted.item() != predicted.item()}")
        print(f"L2 perturbation norm: {l2_norm.item():.4f}")
        print(f"Adversarial confidence: {F.softmax(adv_outputs, dim=1).max().item():.4f}")
        
        # 可视化原始图像和对抗样本
        plt.figure(figsize=(12, 5))
        plt.subplot(1, 3, 1)
        plt.title(f"Original: {predicted.item()}")
        plt.imshow(inputs[0].cpu().permute(1, 2, 0).numpy())
        plt.axis('off')
        
        plt.subplot(1, 3, 2)
        plt.title(f"Adversarial: {adv_predicted.item()}")
        plt.imshow(adv_inputs[0].cpu().permute(1, 2, 0).numpy())
        plt.axis('off')
        
        plt.subplot(1, 3, 3)
        plt.title(f"Difference (x10)")
        diff = (adv_inputs - inputs)[0].cpu().permute(1, 2, 0).numpy()
        # 放大差异使其更容易看到
        plt.imshow((diff * 10 + 0.5).clip(0, 1))
        plt.axis('off')
        
        plt.savefig(f"cw_attack_sample_{labels.item()}.png")
        plt.close()
        
        return adv_predicted.item() != predicted.item(), l2_norm.item()
        
    except Exception as e:
        print(f"Error in CW attack: {e}")
        return False, 0.0

# 在多个样本上测试CW攻击
def test_cw_attack_multiple_samples(model, testloader, num_samples=10):
    success_count = 0
    total_samples = 0
    l2_norms = []
    
    for i, (inputs, labels) in enumerate(testloader):
        if i >= num_samples:
            break
        
        print(f"\n--- Sample {i+1}/{num_samples} ---")
        success, l2_norm = test_cw_attack_on_sample(model, inputs, labels)
        
        if success:
            success_count += 1
            l2_norms.append(l2_norm)
        
        total_samples += 1
    
    success_rate = (success_count / total_samples) * 100
    avg_l2_norm = np.mean(l2_norms) if l2_norms else 0.0
    
    print(f"\nResults on {total_samples} samples:")
    print(f"Success rate: {success_rate:.2f}%")
    print(f"Average L2 norm of successful perturbations: {avg_l2_norm:.4f}")
    
    return success_rate, avg_l2_norm

if __name__ == "__main__":
    print("Loading CIFAR-10 dataset...")
    testloader = load_cifar10()
    
    print("Loading model...")
    model = load_model()
    
    print("Testing CW attack...")
    success_rate, avg_l2_norm = test_cw_attack_multiple_samples(model, testloader, num_samples=5)