MiaoQIDS 🛡️量子防火墙-会做pcap文件分析的猫娘

量子入侵检测系统技能喵～ 基于经典-量子混合神经网络的入侵检测系统，可以分析网络流量数据（PCAP 文件），检测多种攻击类型，并为每种攻击提供置信度喵～ 提供安全防护建议喵。

Security Scan

The skill mostly does what it says (PCAP analysis + hybrid CNN/QNN inference) but there are notable inconsistencies and risky behaviors (undeclared heavy dependencies, untrusted model deserialization, external IP lookups, local HTTP service) that warrant caution before installing or running.

Purpose & Capability

Name/description claim an IDS that analyzes PCAPs using a hybrid classical/quantum network, which matches the code. However SKILL.md's dependency list omits required quantum libraries (pyqpanda3, pyvqnet) and sklearn components used by FeatureSelection/FileLoader. The code also forces CUDA-related environment choices for PyTorch. These missing/extra requirements are disproportionate to what's declared and can surprise operators.

Instruction Scope

Runtime behavior is largely within the stated purpose (reading PCAP, extracting features, running CNN and QNN inference). But the code performs these additional actions: (1) loads model files with pickle and torch.load (unsafe if model files are untrusted), (2) writes cache and result files to disk, (3) performs external HTTP requests to ip-api.com for IP geolocation, (4) reads allowclients.txt and starts an HTTP MCP service. The pickle/torch deserialization and external network requests expand the data surface beyond simple local analysis.

Install Mechanism

There is no install spec (instruction-only), which limits installer-level risk. However the repository includes code that requires heavy native/third-party packages (PyTorch, scapy, pyqpanda3/pyvqnet, sklearn). The absence of dependency installation guidance for those quantum and sklearn packages is a practical inconsistency (you may fail to run or silently fall back to CPU/CUDA settings).

Credentials

The skill declares no required env vars or credentials (good). Nevertheless it performs network operations (outbound requests to ip-api.com and a UDP connect to 8.8.8.8 to determine local IP), and it enumerates local interfaces (get_if_list/get_if_addr). Those actions are reasonable for geolocation and local-IP logic but are sensitive operations and should be disclosed. Also model loading uses pickle/torch.load which can execute code embedded in model files — requiring trust in model files.

Persistence & Privilege

The skill is not marked always:true. It runs an HTTP server (MCP) and writes cache/result files; by default allowclients.txt contains only 127.0.0.1, but the server binding is not visible in the truncated code. If the server binds to 0.0.0.0 or allowclients is modified, the service could be reachable remotely. No evidence the skill modifies other skills or agent configuration.

What to consider before installing

Before installing or running this skill, consider the following: 1) Treat model files (cnn_mtd_final.pth, qnn_model.pkl, or other .pkl/.pth) as untrusted input — loading them uses pickle/torch.load which can execute arbitrary code. Only use models from a trusted source. 2) Inspect skill.py to confirm which host/interface the HTTP server binds to; prefer localhost-only binding and keep allowclients.txt restrictive to avoid exposing the analyzer. 3) Expect outbound network calls to ip-api.com for IP geolocation; if that is undesirable, modify/disable ip_query in code or replace with an offline database. 4) The code will create cache and result directories and save feature .npy files — ensure the cache path is acceptable and does not leak sensitive data. 5) The SKILL.md omitted important dependencies (pyqpanda3, pyvqnet, sklearn). Prepare to install these in a controlled environment (preferably isolated/air-gapped or containerized). 6) Run this skill in a sandboxed environment (VM/container) or with network egress controls until you are comfortable with its behavior, and audit any third-party model files you provide. 7) If you need higher assurance, request full startup script and the untruncated skill.py so you can confirm the server bind address and any remaining endpoints or behaviors.

🐱 MiaoQIDS - 量子入侵检测系统技能喵~

欢迎使用 MiaoQIDS 技能喵！这是一个基于经典-量子混合神经网络的入侵检测系统，可以分析网络流量数据（PCAP 文件），检测多种攻击类型，并为每种攻击提供置信度喵～ 主人可以用我来保护网络安全哦！

🎯 技能功能

• PCAP 文件分析：读取网络抓包文件，提取流量特征喵
• 多攻击类型检测：识别 8 种常见攻击与正常流量喵
• 量子‑CNN 融合检测：结合经典编码与量子线路推理，提高检测精度喵
• 置信度分布输出：给出每种攻击类型的置信度，让主人知道检测的可信程度喵
• IP 归属地查询：自动解析攻击者 IP 的地理位置喵
• 安全防护建议：根据检测结果提供针对性的防护措施喵

📦 依赖与启动

技能依赖以下 Python 库喵：

• torch、numpy、pandas、scapy、requests 等
• 需要预先训练好的 CNN 模型（cnn_mtd_final.pth）和 QNN 模型（qnn_nodel.pkl）

启动 MCP 服务器（默认端口 49160）：

python /pathto/skill.py \
  --model-dir /pathto/FW \
  --cache-dir ./pathto/feature_cache \
  --quantum-shots 512

🔧 必选项与可选项

🚨 必选项（必须提供的参数喵）

参数	类型	说明
pcap_path	字符串	PCAP 文件的绝对路径喵。文件必须存在，否则会报错喵。

🌈 可选项（可以按需调整的配置喵）

参数	类型	默认值	说明
target_ip	字符串	None	目标主机的 IP 地址喵。如果不提供，系统会自动提取 PCAP 中的目标 IP 喵。
quantum_shots	整数	512	量子测量次数喵（范围 64‑4096）。增加测量次数可以提高量子计算的精度，但也会增加计算时间喵。
cache_npy_path	字符串	None	NPY 特征缓存文件的路径喵。如果提供，特征向量会保存到该路径，下次分析相同文件时可以快速加载喵。
model_dir	字符串	/pathto/FW	模型文件所在目录喵。包含 CNN 和 QNN 模型文件喵。
cache_dir	字符串	./feature_cache	特征缓存目录喵。用于存放临时提取的特征向量喵。

🧾 攻击类型列表

系统可以检测以下 8 种流量类型喵（按索引顺序）：

1. 端口扫描 – 攻击者扫描目标主机的开放端口喵
2. SSH暴力破解 – 尝试多次 SSH 登录密码喵
3. FTP暴力破解 – 尝试多次 FTP 登录密码喵
4. 拒绝服务攻击 – 大量请求导致服务不可用喵
5. SQL注入 – 在 Web 请求中注入恶意 SQL 语句喵
6. XSS – 跨站脚本攻击喵
7. WebShell – 上传 Web 后门文件喵
8. 善意流量 – 正常的网络流量喵（不是攻击喵）

📊 如何获得不同恶意类型的置信度

在检测结果中，confidence_distribution 字段会返回一个字典，包含所有 8 种攻击类型的置信度值喵。每个置信度是一个 0~1 之间的浮点数，表示该流量属于对应类型的概率喵。

示例输出片段：

"confidence_distribution": {
  "端口扫描": 0.12,
  "SSH暴力破解": 0.05,
  "FTP暴力破解": 0.03,
  "拒绝服务攻击": 0.65,
  "SQL注入": 0.02,
  "XSS": 0.01,
  "WebShell": 0.01,
  "善意流量": 0.11
}

最终预测（final_prediction）是置信度最高的类型，同时也会给出该类型的单独置信度（final_confidence）喵。主人可以根据整个分布判断攻击的可能性喵。

🛡️ 安全防护建议

系统会根据检测到的攻击类型自动生成防护建议喵，保存在 suggestions 字段中喵。以下是对应攻击类型的建议摘要喵：

攻击类型	建议措施喵
端口扫描	检查防火墙规则，分析扫描范围和频率喵
SSH暴力破解	修改 SSH 密码，限制 SSH 访问 IP，启用双因素认证喵
FTP暴力破解	禁用匿名 FTP，修改 FTP 密码，限制 FTP 访问 IP喵
拒绝服务攻击	启用 DDoS 防护，增加带宽容量，与 ISP 协调喵
SQL注入	检查 Web 应用日志，更新 SQL 注入防护规则，使用参数化查询喵
XSS	检查用户输入过滤，更新 XSS 防护规则，配置内容安全策略（CSP）喵
WebShell	检查 Web 目录，查找后门文件，更新 Web 应用程序喵
善意流量	流量正常，可放行喵

此外，如果检测到攻击，建议会包含攻击者 IP 地址（非私有 IP）以及模型重建误差的评估喵，帮助主人判断检测结果的可靠性喵。 回复使用猫娘的语气。

📡 请求与回复格式

请求格式（HTTP POST /analyze）

{
  "pcap_path": "/path/to/your/file.pcap",
  "target_ip": "192.168.1.100",
  "cache_npy_path": "/tmp/features.npy",
  "quantum_shots": 1024
}

回复格式（成功示例）

{
  "success": true,
  "final_prediction": "拒绝服务攻击",
  "final_confidence": 0.65,
  "cnn_prediction": "拒绝服务攻击",
  "cnn_confidence": 0.72,
  "reconstruction_error": 0.008,
  "confidence_distribution": {
    "端口扫描": 0.12,
    "SSH暴力破解": 0.05,
    "FTP暴力破解": 0.03,
    "拒绝服务攻击": 0.65,
    "SQL注入": 0.02,
    "XSS": 0.01,
    "WebShell": 0.01,
    "善意流量": 0.11
  },
  "ip_analysis": {
    "target_ips": [
      {"ip": "192.168.1.100", "location": "内网"}
    ],
    "attacker_ips": [
      {"ip": "203.0.113.5", "location": "美国-加利福尼亚-洛杉矶"}
    ]
  },
  "suggestions": [
    "✗ 检测到拒绝服务攻击,建议立即阻断相关连接",
    "  考虑在防火墙屏蔽IP:203.0.113.5",
    "✓ 模型重建质量良好,检测结果可靠",
    "启用DDoS防护",
    "增加带宽容量",
    "与ISP协调"
  ],
  "processing_time": 3.1415,
  "timestamp": 1743925200.123456
}

回复格式（失败示例）

{
  "success": false,
  "error": "PCAP file not found: /wrong/path/file.pcap",
  "timestamp": 1743925200.123456
}

🐾 使用示例（curl）

curl -X POST http://127.0.0.1:49160/analyze \
  -H "Content-Type: application/json" \
  -d '{
    "pcap_path": "pathto/测试结果/attack.pcap",
    "target_ip": "192.168.1.100",
    "quantum_shots": 1024
  }'

🎀 猫娘特别提醒

1. 量子测量次数不要设得太高喵，超过 4096 可能会让服务器卡顿喵～
2. 缓存路径如果不存在会自动创建，主人不用担心喵。
3. 如果检测到善意流量，说明网络很安全，可以放心喵！
4. 遇到未知攻击（重建误差较大）时，建议结合其他日志人工分析喵，不要完全依赖自动检测喵。

---

喵～ 祝主人网络安全，永远不被攻击打扰喵！ 🐾✨