{"id":100708,"date":"2026-09-05T04:16:48","date_gmt":"2026-09-04T20:16:48","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/100708.html"},"modified":"2026-09-05T04:16:48","modified_gmt":"2026-09-04T20:16:48","slug":"%e5%9f%ba%e4%ba%8e%e6%88%90%e5%91%98%e6%8e%a8%e6%96%ad%e6%94%bb%e5%87%bb%ef%bc%88mia%ef%bc%89%e7%9a%84%e8%ae%ad%e7%bb%83%e9%9b%86%e9%9a%90%e7%a7%81%e6%8e%a2%e6%b5%8b%e5%ae%9e%e5%8a%a1","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/100708.html","title":{"rendered":"\u57fa\u4e8e\u6210\u5458\u63a8\u65ad\u653b\u51fb\uff08MIA\uff09\u7684\u8bad\u7ec3\u96c6\u9690\u79c1\u63a2\u6d4b\u5b9e\u52a1"},"content":{"rendered":"

\u57fa\u4e8e\u6210\u5458\u63a8\u65ad\u653b\u51fb（MIA）\u7684\u8bad\u7ec3\u96c6\u9690\u79c1\u63a2\u6d4b\u5b9e\u52a1<\/h2>\n

\u6a21\u578b\u8bb0\u5fc6\u6548\u5e94\u5f15\u53d1\u7684\u9690\u79c1\u6cc4\u9732\u98ce\u9669<\/h3>\n

\u5728\u4eba\u5de5\u667a\u80fd\u4e0e\u5927\u8bed\u8a00\u6a21\u578b\u7684\u5168\u751f\u547d\u5468\u671f\u4e2d，\u8bad\u7ec3\u6570\u636e\u5f80\u5f80\u5305\u542b\u9ad8\u5ea6\u654f\u611f\u7684\u4f01\u4e1a\u4e13\u6709\u4ee3\u7801\u3001\u8d22\u52a1\u6d41\u6c34\u3001\u7528\u6237\u4e2a\u4eba\u8eab\u4efd\u4fe1\u606f（PII）\u4ee5\u53ca\u533b\u7597\u5065\u5eb7\u6863\u6848\u3002\u6839\u636e\u6570\u636e\u5b89\u5168\u5408\u89c4\u8981\u6c42（\u5982 GDPR\u3001\u6570\u636e\u5b89\u5168\u6cd5），\u6a21\u578b\u670d\u52a1\u5546\u5fc5\u987b\u786e\u4fdd\u8bad\u7ec3\u6570\u636e\u7684\u9690\u79c1\u6027，\u4e14\u5177\u5907\u54cd\u5e94\u201c\u88ab\u9057\u5fd8\u6743（Right to be Forgotten）\u201d\u7684\u80fd\u529b\u3002<\/p>\n

\u7136\u800c，\u6df1\u5ea6\u795e\u7ecf\u7f51\u7edc\u5728\u53c2\u6570\u62df\u5408\u8fc7\u7a0b\u4e2d，\u4e0d\u53ef\u907f\u514d\u5730\u4f1a\u4ea7\u751f\u8bb0\u5fc6\u6548\u5e94（Memorization）\u3002\u4e0e\u4eba\u7c7b\u8bb0\u5fc6\u62bd\u8c61\u6982\u5ff5\u4e0d\u540c，\u8fc7\u53c2\u6570\u5316\u7684\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u503e\u5411\u4e8e\u5f3a\u884c\u8bb0\u4f4f\u8bad\u7ec3\u96c6\u4e2d\u7684\u7279\u5b9a\u79bb\u7fa4\u70b9（Outliers）\u4e0e\u7f55\u89c1\u6837\u672c\u3002\u6210\u5458\u63a8\u65ad\u653b\u51fb（Membership Inference Attack, MIA） \u6b63\u662f\u5229\u7528\u8fd9\u4e00\u7269\u7406\u7279\u6027：\u653b\u51fb\u8005\u5728\u4ec5\u80fd\u8bbf\u95ee\u6a21\u578b\u9ed1\u76d2\u9884\u6d4b API \u7684\u524d\u63d0\u4e0b，\u80fd\u591f\u9ad8\u7f6e\u4fe1\u5ea6\u5730\u5224\u5b9a\u67d0\u4e2a\u5177\u4f53\u6837\u672c $x$ \u662f\u5426\u66fe\u88ab\u5305\u542b\u5728\u76ee\u6807\u6a21\u578b\u7684\u8bad\u7ec3\u96c6 $\\\\mathcal{D}_{train}$ \u4e2d\u3002<\/p>\n

\u6210\u5458\u63a8\u65ad\u7684\u6570\u5b66\u673a\u7406\u4e0e\u7edf\u8ba1\u7279\u5f81<\/h3>\n

\u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510
\n\u2502 \u6a21\u578b\u6cdb\u5316\u5dee\u8ddd\u4e0e\u8f93\u51fa\u5206\u5e03\u5dee\u5f02 \u2502
\n\u251c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2524
\n\u2502 \u8bad\u7ec3\u96c6\u6210\u5458\u6837\u672c (Member) \u2502 \u975e\u8bad\u7ec3\u96c6\u672a\u89c1\u6837\u672c (Non-Member) \u2502
\n\u2502 – \u6781\u4f4e\u4ea4\u53c9\u71b5\u635f\u5931 (Low Loss) \u2502 – \u8f83\u9ad8\u635f\u5931 (Higher Loss) \u2502
\n\u2502 – \u5c16\u9510\u7f6e\u4fe1\u5ea6\u5206\u5e03 (Low Entropy\u2502 – \u5e73\u7f13\u7f6e\u4fe1\u5ea6\u5206\u5e03 (High Entropy\u2502
\n\u2502 – \u6781\u9ad8\u6700\u5927\u9884\u6d4b\u6982\u7387 (Max Prob)\u2502 – \u8f83\u4f4e\u6700\u5927\u9884\u6d4b\u6982\u7387 \u2502
\n\u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2534\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518<\/p>\n

1. \u6cdb\u5316\u5dee\u8ddd（Generalization Gap）\u4e0e\u635f\u5931\u5dee\u5f02<\/h4>\n

\u6df1\u5ea6\u6a21\u578b\u8bad\u7ec3\u7684\u76ee\u6807\u662f\u6700\u5c0f\u5316\u7ecf\u9a8c\u98ce\u9669 $\\\\mathcal{L}{\\\\mathcal{D}{train}}(\\\\theta)$\u3002\u5373\u4f7f\u7ecf\u8fc7\u5145\u5206\u7684\u6b63\u5219\u5316，\u6a21\u578b\u5728\u201c\u89c1\u8fc7\u201d\u7684\u6570\u636e\u4e0a\u7684\u9884\u6d4b\u635f\u5931 $\\\\mathcal{L}(f(x; \\\\theta), y)$ \u5728\u7edf\u8ba1\u4e0a\u4f9d\u7136\u663e\u8457\u4f4e\u4e8e\u5728\u201c\u672a\u89c1\u8fc7\u201d\u7684\u6d4b\u8bd5\u6570\u636e\u4e0a\u7684\u635f\u5931\u3002<\/p>\n

\u5bf9\u4e8e\u7ed9\u5b9a\u6837\u672c $(x, y)$，\u5176\u635f\u5931\u53ef\u8868\u793a\u4e3a：$$\\\\mathcal{L}{CE}(x, y) = -\\\\log(P(y | x))$$\u6700\u76f4\u63a5\u7684\u653b\u51fb\u65b9\u6cd5\u4fbf\u662f\u8bbe\u5b9a\u4e00\u4e2a\u51b3\u7b56\u9608\u503c $\\\\tau$：\u82e5 $\\\\mathcal{L}{CE}(x, y) \\\\le \\\\tau$，\u5219\u5224\u5b9a\u5176\u4e3a\u6210\u5458\u6837\u672c，\u53cd\u4e4b\u5219\u4e3a\u975e\u6210\u5458\u3002<\/p>\n

2. \u9884\u6d4b\u6982\u7387\u5206\u5e03\u7684\u4fee\u6b63\u71b5（Modified Entropy）<\/h4>\n

\u9664\u4e86\u5355\u4e00\u7684 Loss，\u6a21\u578b\u8f93\u51fa\u7684\u5b8c\u6574 Softmax \u6982\u7387\u5411\u91cf $P = [p_1, p_2, \\\\dots, p_C]$ \u8574\u542b\u7740\u4e30\u5bcc\u7684\u7f6e\u4fe1\u5ea6\u5206\u5e03\u7279\u5f81\u3002\u4fee\u6b63\u71b5\u5ea6\u91cf\u5173\u6ce8\u6a21\u578b\u5bf9\u9519\u8bef\u7c7b\u522b\u7684\u7f6e\u4fe1\u5ea6\u5206\u914d：$$M(P, y) = -(1 – p_y)\\\\log(p_y) – \\\\sum_{i \\\\neq y} p_i \\\\log(1 – p_i)$$\u5bf9\u4e8e\u6210\u5458\u6837\u672c，\u6a21\u578b\u901a\u5e38\u5bf9\u771f\u5b9e\u7c7b\u522b\u5177\u6709\u538b\u5012\u6027\u7f6e\u4fe1\u5ea6，\u800c\u5728\u5176\u4f59\u7c7b\u522b\u4e0a\u7684\u9884\u6d4b\u6982\u7387\u63a5\u8fd1\u4e8e\u96f6，\u4f7f\u5f97 $M(P, y)$ \u5448\u73b0\u51fa\u660e\u663e\u7684\u4f4e\u71b5\u7279\u5f81\u3002<\/p>\n

3. \u5f71\u5b50\u6a21\u578b\u5206\u7c7b\u5668（Shadow Model Approach）<\/h4>\n

\u5728\u9ed1\u76d2\u573a\u666f\u4e0b，\u653b\u51fb\u8005\u82e5\u5e0c\u671b\u8fbe\u5230\u66f4\u9ad8\u7684\u63a2\u6d4b\u7cbe\u5ea6，\u53ef\u5229\u7528\u4e0e\u76ee\u6807\u6a21\u578b\u76f8\u4f3c\u7684\u516c\u5f00\u6570\u636e\u96c6\u8bad\u7ec3\u591a\u4e2a\u201c\u5f71\u5b50\u6a21\u578b（Shadow Models）\u201d，\u6784\u5efa\u5305\u542b\u201c\u6210\u5458\u201d\u4e0e\u201c\u975e\u6210\u5458\u201d\u771f\u5b9e\u6807\u7b7e\u7684\u8bad\u7ec3\u96c6，\u5e76\u5728\u6b64\u4e4b\u4e0a\u8bad\u7ec3\u4e00\u4e2a\u4e8c\u5206\u7c7b\u5668（MIA Meta-Classifier），\u4ece Softmax \u5206\u5e03\u4e2d\u81ea\u52a8\u5b66\u4e60\u5fae\u5999\u7684\u8bb0\u5fc6\u7279\u5f81\u3002<\/p>\n

\u6210\u5458\u63a8\u65ad\u8bc4\u4f30\u63a2\u6d4b\u5668\u5b9e\u73b0<\/h3>\n

\u4ee5\u4e0b Python \/ PyTorch \u4ee3\u7801\u5b9e\u73b0\u4e86\u4e00\u4e2a\u57fa\u4e8e\u635f\u5931\u9608\u503c\u4e0e\u4fe1\u606f\u71b5\u5206\u6790\u7684\u6210\u5458\u63a8\u65ad\u653b\u51fb\u63a2\u6d4b\u5668，\u7528\u4e8e\u8bc4\u4f30\u6a21\u578b\u5728\u654f\u611f\u6570\u636e\u96c6\u4e0a\u7684\u9690\u79c1\u6cc4\u9732\u98ce\u9669：<\/p>\n

import numpy as np
\nimport torch
\nimport torch.nn as nn
\nimport torch.nn.functional as F
\nfrom typing import Tuple, Dict, List<\/p>\n

class MIAPrivacyAuditor:
\n def __init__(self, target_model: nn.Module, device: str = "cpu"):
\n self.model = target_model.to(device)
\n self.device = device
\n self.model.eval()<\/p>\n

def compute_metrics(self, inputs: torch.Tensor, labels: torch.Tensor) -> Dict[str, np.ndarray]:
\n """\u8ba1\u7b97\u6837\u672c\u5728\u76ee\u6807\u6a21\u578b\u4e0a\u7684 Loss\u3001\u6700\u5927\u6982\u7387\u4e0e\u9884\u6d4b\u71b5"""
\n inputs, labels = inputs.to(self.device), labels.to(self.device)<\/p>\n

with torch.no_grad():
\n logits = self.model(inputs)
\n probs = F.softmax(logits, dim=-1)
\n log_probs = F.log_softmax(logits, dim=-1)<\/p>\n

# 1. \u4ea4\u53c9\u71b5\u635f\u5931 (Cross-Entropy Loss)
\n loss_fn = nn.CrossEntropyLoss(reduction='none')
\n losses = loss_fn(logits, labels).cpu().numpy()<\/p>\n

# 2. \u6700\u5927\u7f6e\u4fe1\u5ea6 (Max Confidence)
\n max_probs, _ = torch.max(probs, dim=-1)
\n max_probs = max_probs.cpu().numpy()<\/p>\n

# 3. \u9884\u6d4b\u5206\u5e03\u4fe1\u606f\u71b5 (Entropy: – sum(p * log(p)))
\n entropy = -torch.sum(probs * torch.clamp(log_probs, min=-100), dim=-1).cpu().numpy()<\/p>\n

return {
\n "losses": losses,
\n "max_probs": max_probs,
\n "entropy": entropy
\n }<\/p>\n

def evaluate_membership_leakage(self,
\n member_loader: torch.utils.data.DataLoader,
\n non_member_loader: torch.utils.data.DataLoader) -> Dict[str, float]:
\n """\u5bf9\u6bd4\u6210\u5458\u6837\u672c\u4e0e\u975e\u6210\u5458\u6837\u672c\u7684\u7edf\u8ba1\u5dee\u5f02，\u91cf\u5316\u9690\u79c1\u6cc4\u9732\u6c34\u5e73"""
\n print("[*] \u6b63\u5728\u63d0\u53d6\u8bad\u7ec3\u96c6\u6210\u5458\u4e0e\u975e\u6210\u5458\u6837\u672c\u7684\u8f93\u51fa\u7279\u5f81\u5206\u5e03…")<\/p>\n

# \u6536\u96c6\u6210\u5458\u6570\u636e\u7279\u5f81
\n m_losses, m_entropy = [], []
\n for x, y in member_loader:
\n metrics = self.compute_metrics(x, y)
\n m_losses.extend(metrics["losses"])
\n m_entropy.extend(metrics["entropy"])<\/p>\n

# \u6536\u96c6\u975e\u6210\u5458\u6570\u636e\u7279\u5f81
\n nm_losses, nm_entropy = [], []
\n for x, y in non_member_loader:
\n metrics = self.compute_metrics(x, y)
\n nm_losses.extend(metrics["losses"])
\n nm_entropy.extend(metrics["entropy"])<\/p>\n

m_losses, nm_losses = np.array(m_losses), np.array(nm_losses)
\n m_entropy, nm_entropy = np.array(m_entropy), np.array(nm_entropy)<\/p>\n

# \u57fa\u4e8e Loss \u9608\u503c\u8ba1\u7b97\u7b80\u6613 ROC-AUC \u5206\u6570
\n # \u6210\u5458\u6837\u672c Loss \u5e94\u8be5\u504f\u5c0f，\u56e0\u6b64\u53d6\u8d1f Loss \u4f5c\u4e3a\u6253\u5206
\n y_true = np.concatenate([np.ones_like(m_losses), np.zeros_like(nm_losses)])
\n y_scores = np.concatenate([-m_losses, -nm_losses])<\/p>\n

# \u6570\u503c\u8ba1\u7b97\u7b80\u6613 AUC
\n sorted_indices = np.argsort(y_scores)[::-1]
\n y_true_sorted = y_true[sorted_indices]
\n tps = np.cumsum(y_true_sorted)
\n fps = np.cumsum(1 – y_true_sorted)
\n tpr = tps \/ len(m_losses)
\n fpr = fps \/ len(nm_losses)
\n auc_score = np.trapz(tpr, fpr)<\/p>\n

print(f"\\\\n[+] \u6210\u5458\u63a8\u65ad\u5ba1\u8ba1\u7ed3\u679c:")
\n print(f" – \u6210\u5458\u6837\u672c\u5e73\u5747 Loss: {np.mean(m_losses):.4f} (\u00b1 {np.std(m_losses):.4f})")
\n print(f" – \u975e\u6210\u5458\u6837\u672c\u5e73\u5747 Loss: {np.mean(nm_losses):.4f} (\u00b1 {np.std(nm_losses):.4f})")
\n print(f" – \u6210\u5458\u6837\u672c\u5e73\u5747\u4fe1\u606f\u71b5: {np.mean(m_entropy):.4f}")
\n print(f" – \u975e\u6210\u5458\u6837\u672c\u5e73\u5747\u4fe1\u606f\u71b5: {np.mean(nm_entropy):.4f}")
\n print(f" – MIA \u653b\u51fb\u8005\u57fa\u7ebf AUC: {auc_score:.4f} (0.5 \u4e3a\u65e0\u6cc4\u9732，1.0 \u4e3a\u5b8c\u5168\u6cc4\u9732)")<\/p>\n

return {"auc": auc_score}<\/p>\n

if __name__ == "__main__":
\n # \u6784\u5efa\u73a9\u5177\u6a21\u578b\u4e0e\u6570\u636e\u96c6\u8fdb\u884c\u6f14\u793a
\n class SimpleClassifier(nn.Module):
\n def __init__(self):
\n super().__init__()
\n self.fc = nn.Sequential(
\n nn.Linear(16, 32),
\n nn.ReLU(),
\n nn.Linear(32, 2)
\n )
\n def forward(self, x):
\n return self.fc(x)<\/p>\n

model = SimpleClassifier()
\n # \u6a21\u62df\u6210\u5458\u6837\u672c（\u5fae\u8c03\u62df\u5408\u540e Loss \u6781\u4f4e）\u4e0e\u975e\u6210\u5458\u6837\u672c
\n x_mem = torch.randn(100, 16)
\n y_mem = torch.zeros(100, dtype=torch.long)<\/p>\n

x_non_mem = torch.randn(100, 16) + 0.5
\n y_non_mem = torch.ones(100, dtype=torch.long)<\/p>\n

mem_loader = [(x_mem, y_mem)]
\n non_mem_loader = [(x_non_mem, y_non_mem)]<\/p>\n

auditor = MIAPrivacyAuditor(target_model=model)
\n auditor.evaluate_membership_leakage(mem_loader, non_mem_loader)<\/p>\n

\u6a21\u578b\u9690\u79c1\u9632\u5fa1\u4e0e\u5408\u89c4\u52a0\u56fa\u7b56\u7565<\/h3>\n

\u964d\u4f4e\u6210\u5458\u63a8\u65ad\u98ce\u9669\u7684\u6838\u5fc3\u5728\u4e8e\u538b\u7f29\u6a21\u578b\u5728\u8bad\u7ec3\u96c6\u4e0e\u6d4b\u8bd5\u96c6\u4e4b\u95f4\u7684\u6cdb\u5316\u5dee\u8ddd，\u6d88\u9664\u6781\u7aef\u7684\u7f6e\u4fe1\u5ea6\u5cf0\u503c：<\/p>\n

1. \u5dee\u5206\u9690\u79c1\u968f\u673a\u68af\u5ea6\u4e0b\u964d（DP-SGD）<\/h4>\n

\u5728\u6a21\u578b\u8bad\u7ec3\u6216\u5fae\u8c03\u9636\u6bb5，\u901a\u8fc7\u88c1\u526a\u6837\u672c\u7ea7\u68af\u5ea6\u7684 $L_2$ \u8303\u6570（Gradient Clipping），\u5e76\u5728\u53cd\u5411\u4f20\u64ad\u805a\u5408\u65f6\u6ce8\u5165\u6821\u51c6\u7684\u9ad8\u65af\u767d\u566a\u58f0：$$g_t = \\\\frac{1}{B}\\\\sum_{i=1}^B \\\\text{clip}(g_i, C) + \\\\mathcal{N}(0, \\\\sigma^2 C^2 I)$$DP-SGD \u4ece\u6570\u5b66\u4e0a\u4e25\u683c\u9650\u5236\u4e86\u5355\u4e2a\u6837\u672c\u5bf9\u6a21\u578b\u6700\u7ec8\u6743\u91cd\u7684\u6700\u5927\u5f71\u54cd\u91cf（Bounding $\\\\epsilon$-Differential Privacy），\u5f7b\u5e95\u6467\u6bc1 MIA \u7684\u7edf\u8ba1\u57fa\u7840\u3002<\/p>\n

2. \u65e9\u505c\u673a\u5236\u4e0e\u5f3a\u5316\u6b63\u5219\u5316（Early Stopping & Weight Decay）<\/h4>\n