[教學] 掃描模型的簡單範例

註記：此頁為由 AI（gpt-5-mini-2025-08-07）自動翻譯自英文原文，可能含有少量不準確之處。

📌 介紹

本文示範如何使用一個簡單的範例與 modelscan 工具來偵測不安全的 PyTorch 模型。

🚀 快速開始

開始之前，您需要安裝以下套件

pip install numpy torch modelscan

準備安全模型

from torch import nn
import torch

class SafeModel(nn.Module):

    def __init__(self):
        super(SafeModel, self).__init__()
        self.linear = nn.Linear(10, 1)

    def forward(self, x):
        return self.linear(x)
    
if __name__ == "__main__":
    model = SafeModel()

    # save the model
    torch.save(model.state_dict(), "safe_model.pth")

準備惡意模型

這是一個在載入時會產生輸出的惡意模型。

from torch import nn
import torch
import os

class MaliciousModel:

    def __reduce__(self):
        print("Reduce called!")  # 應該會印出
        return (os.system, ("echo 'This is a malicious model!' > malicious_output.txt",))
    
if __name__ == "__main__":
    model = MaliciousModel()

    # save the model
    torch.save(model, "malicious_model.pth")

載入模型

Torch 已經有基本保護，因此我們需要暫時關閉 weights_only 選項。當你載入該模型後，你會看到名為 malicious_output.txt 的檔案。這表示惡意行為已經在載入時發生。

import torch

safe_model_path = "safe_model.pth"
malicious_model_path = "malicious_model.pth"

s_model = torch.load(safe_model_path)
m_model = torch.load(malicious_model_path, weights_only=False)

使用 modelscan 掃描模型

安全模型

modelscan -p safe_model.pth

Scanning /Users/hsiangjenli/Documents/github/mlops-survey/safe_model.pth:safe_model/data.pkl using modelscan.scanners.PickleUnsafeOpScan model scan

--- Summary ---

 No issues found! 🎉

--- Skipped --- 

Total skipped: 7 - run with --show-skipped to see the full list.

惡意模型

modelscan -p malicious_model.pth

Scanning /Users/hsiangjenli/Documents/github/mlops-survey/malicious_model.pth:malicious_model/data.pkl using modelscan.scanners.PickleUnsafeOpScan model scan

--- Summary ---

Total Issues: 1

Total Issues By Severity:

    - LOW: 0
    - MEDIUM: 0
    - HIGH: 0
    - CRITICAL: 1

--- Issues by Severity ---

--- CRITICAL ---

Unsafe operator found:
  - Severity: CRITICAL
  - Description: Use of unsafe operator 'system' from module 'posix'
  - Source: /Users/hsiangjenli/Documents/github/mlops-survey/malicious_model.pth:malicious_model/data.pkl

--- Skipped --- 

Total skipped: 5 - run with --show-skipped to see the full list.

🔁 重點回顧

1. 建立了一個安全模型與一個惡意模型（在載入時會產生輸出）
2. 使用 modelscan 掃描了兩個模型

🔗 參考資料

• https://github.com/protectai/modelscan