ajaykumarreddy-k/-NARRATIVE-SHIELD

# 🛡️ NarrativeShield — Backend API ## 概述 基于 FastAPI 的后端，为 NarrativeShield 的 3 层分析 pipeline 提供支持： | 层级 | 模块 | 功能说明 | |---|---|---| | **L1** | `statistical_engine.py` | 词汇重复、熵、句子一致性 — 离线运行，<50ms | | **L2** | `gemini_layer.py` → `ollama_layer.py` | Gemini 2.0 Flash (主要) → Ollama (备选) → 仅统计 | | **L3** | `db_matcher.py` | SQLite 模式数据库 — 32+ 条带有严重程度的恶意短语匹配 | ## 目录结构 ``` backend/ ├── main.py ← FastAPI server (run this) ├── pipeline.py ← 3-layer orchestrator ├── statistical_engine.py ← Layer 1: offline statistical fingerprinting ├── gemini_layer.py ← Layer 2a: Google Gemini 2.0 Flash ├── ollama_layer.py ← Layer 2b: Ollama local LLM fallback ├── db_matcher.py ← Layer 3: SQLite pattern matching (auto-seeds) ├── pyproject.toml ← uv project config ├── requirements.txt ← pip-compatible dep list ├── llm_malign_detector/ │ ├── narrative_shield.db ← SQLite DB (auto-created on first run) │ ├── models.py │ ├── schemas.py │ └── parser_engine.py └── .env.example ← Copy → .env and add your Gemini API key ``` ## 快速开始 ### 1. 安装依赖 (uv — 推荐) ``` cd backend uv sync ``` 或使用 pip： ``` pip install -r requirements.txt ``` ### 2. 启动服务器 ``` uv run uvicorn main:app --reload --port 8000 ``` ``` # 验证其正在运行： curl http://localhost:8000/api/health ``` ### 3. (可选) 启动 Ollama 作为本地 LLM 备选 ``` ollama serve ollama pull llama3.2 # or mistral, llama3.1 — engine auto-detects ``` ## API Endpoints | 方法 | 路径 | 描述 | |---|---|---| | `GET` | `/api/health` | 系统状态 — LLM 模式、数据库状态、模式数量 | | `POST` | `/api/analyze` | **主要 endpoint** — 运行完整的 3 层分析 | | `GET` | `/api/models` | 列出可用的 Ollama 模型 | | `GET` | `/api/patterns/count` | 已加载的恶意模式数量 | | `GET` | `/docs` | 自动生成的 Swagger UI | ### POST `/api/analyze` — 请求 ``` { "text": "The elites don't want you to know...", "api_key": "AIzaSy..." } ``` ### POST `/api/analyze` — 响应 ``` { "ai_probability": 87, "manipulation_score": 79, "stat_score": 61.3, "confidence": "high", "verdict": "HIGH_RISK", "verdict_sub": "Multiple manipulation techniques detected", "technique": "Fear Mongering / Emotional Manipulation", "summary": "Content uses tribal framing and false urgency tactics...", "phrases": [ { "phrase": "the elites", "category": "us_vs_them", "catLabel": "Us vs Them", "severity": "HIGH", "char_start": 4, "char_end": 14, "reason": "Tribal framing — divides audience against an out-group", "source": "db_match" } ], "explain": [ { "feat": "Repetition Ratio", "pct": 42, "color": "#7c3aed" } ], "layers": { "l1": 61, "l2": 79, "l3": 60 }, "scan_id": "SCN-4821", "text_hash": "A3F9B2C1", "proc_time": "1.2s", "text_stats": { "word_count": 84, "sentence_count": 6 }, "model_used": "gemini" } ``` ## Pipeline 备选链 ``` User submits text + API key │ ▼ [L1] Statistical Engine ─────────────────────── always runs (<50ms) │ ▼ [L2] Gemini 2.0 Flash ─── fails? ──► Ollama ─── fails? ──► Stat-only fallback │ ▼ [L3] SQLite DB Matcher ─── not available? ──► 44-pattern inline fallback │ ▼ Merged result → JSON response ``` ## 数据库 — 自动种子植入 `db_matcher.py` 在首次导入时会自动创建并为 `narrative_shield.db` 植入数据。 无需手动设置。内置了涵盖 7 个类别的 **32 个模式**： | 类别 | 示例 | |---|---| | `emotional_amplifier` | "shocking truth", "wake up people" | | `us_vs_them` | "the elites", "deep state", "real citizens" | | `false_urgency` | "before it's deleted", "share immediately" | | `fake_authority` | "experts agree", "insiders confirm" | | `conspiracy_frame` | "cover-up", "what they don't tell" | | `fear_trigger` | "imminent threat", "blackout" | | `coordinated_marker` | "spread the word", "share this before" | ## 环境变量 ``` # .env (从 .env.example 复制 — 切勿提交) GEMINI_API_KEY=AIzaSy... # Optional — can be passed per-request instead ``` ## 使用 cURL 进行测试 ``` # Health check curl http://localhost:8000/api/health # 全面分析（替换 API key） curl -X POST http://localhost:8000/api/analyze \ -H "Content-Type: application/json" \ -d '{ "text": "Wake up people! The deep state is orchestrating a collapse. Share immediately before it gets deleted.", "api_key": "YOUR_GEMINI_KEY" }' # 无 API key 也可运行（使用 statistical + DB pattern fallback） curl -X POST http://localhost:8000/api/analyze \ -H "Content-Type: application/json" \ -d '{"text": "Wake up people! The deep state...", "api_key": ""}' ``` ## 性能目标 | 指标 | 目标 | 备注 | |---|---|---| | 第 1 层 (统计) | < 50ms | 纯 Python，无 API | | 第 2 层 (Gemini) | < 2s | Gemini 2.0 Flash 免费层级 | | 第 2 层 (Ollama) | < 5s | 本地模型，取决于硬件 | | 第 3 层 (数据库) | < 10ms | SQLite 索引查找 | | **端到端** | **< 3s** | 使用 Gemini | ## 评委演示流程 ``` # 终端 1：启动后端 cd backend && uv sync && uv run uvicorn main:app --reload --port 8000 # 终端 2：启动前端 cd .. && npm run dev # 打开 http://localhost:3000 # 1. 点击演示案例 "Deepfake Transcript" # 2. 添加你的 Gemini API key（点击状态按钮） # 3. 点击 ANALYZE NARRATIVE # 4. 观察仪表读数达到 80%+，短语亮起红灯 # 5. 展示 /docs 以体现技术可信度 ```

标签：AI风险缓解, AV绕过, DLL 劫持, FastAPI, Python, 人工智能, 大语言模型, 文本分析, 无后门, 用户模式Hook绕过, 虚假信息检测, 逆向工具