anti-fraud（反欺诈）

Anti-Fraud & Bot Detection System

A three-layer defense system for registration forms that detects bots while minimizing false positives for legitimate users.

Architecture Overview

Layer 1: SERVER-SIDE (tamper-proof)
├── Encrypted timestamp token verification
├── Device fingerprint matching
└── Minimum fill time enforcement (3s)

Layer 2: MANIPULATION DETECTION (server comparison)
├── Client vs Server timing mismatch
├── Keystroke/input inconsistency
└── Impossible value detection

Layer 3: CLIENT SIGNALS (informational)
├── Honeypot fields
├── Behavioral analysis
└── Content analysis

Key Principle: Never trust client-side data alone.

Quick Implementation

1. Form Token Endpoint

// /api/auth/form-token
// Generate AES-256-GCM encrypted token with timestamp
const token = encrypt({ timestamp: Date.now(), fingerprint, nonce });

2. Behavior Tracking Hook

interface BehaviorSignals {
  totalFillTimeMs: number;
  fieldTimings: Record<string, number>;
  inputMethods: Record<string, 'typed' | 'pasted' | 'autofilled' | 'mixed'>;
  keystrokes: KeystrokeData[];
  keystrokeVariance: number;
  mouseMovements: MouseMovement[];
  hasMouseActivity: boolean;
  focusSequence: string[];
  tabKeyUsed: boolean;
  backspaceCount: number;
}

3. Honeypot Fields

Add hidden fields (CSS hidden, aria-hidden, tabIndex=-1):

• website, phone2, address, company

Any content in honeypot → Instant shadow ban

Risk Scoring

Shadow Ban Triggers (ANY = ban)

Signal Weights

See references/signal-weights.md for complete weight tables.

Critical (+100): HONEYPOT_FILLED, DISPOSABLE_EMAIL High (+25-40): INSTANT_SUBMIT, ALL_FIELDS_PASTED, BOT_PASSWORD_PATTERN, NO_MOUSE_MOVEMENT Positive (-5 to -40): PASSWORD_MANAGER_LIKELY, KEYBOARD_ONLY_USER, NATURAL_TYPING_RHYTHM

Shadow Ban Response

if (shouldShadowBan) {
  await delay(1000 + Math.random() * 2000); // Appear legitimate
  return Response.json({ message: 'Registration successful' }, { status: 200 });
  // No account created, no backend call
}

False Positive Prevention

Password Manager Detection (-40 points)

const isPasswordManager =
  allFieldsAutofilledOrPasted &&
  keystrokeCount < 5 &&
  fillTime >= 1000 && fillTime < 15000;

Keyboard-Only User Detection (-15 points)

const isKeyboardOnly =
  tabKeyUsed &&
  focusSequence.length >= 2 &&
  !hasMouseActivity &&
  totalFieldTime > 1000;

File Structure

src/
├── lib/anti-fraud/
│   ├── index.ts
│   ├── types.ts
│   ├── constants.ts
│   ├── risk-scoring.ts
│   ├── server-token.ts
│   ├── manipulation-detector.ts
│   └── validators/
│       ├── email-validator.ts
│       ├── name-validator.ts
│       └── password-validator.ts
├── hooks/use-behavior-tracking.ts
├── components/anti-fraud/honeypot-fields.tsx
└── app/api/auth/
    ├── form-token/route.ts
    └── register/route.ts

Resources

• Signal weights & thresholds: See references/signal-weights.md
• Validators (email, name, password): See references/validators.md
• XML patterns & detection: See references/detection-patterns.md

Environment

AUTH_SECRET=your-secret-key-for-token-encryption

Logging

All decisions logged with [ANTI_FRAUD] prefix:

[ANTI_FRAUD] { timestamp, emailDomain, serverFillTimeMs, summary: 'Risk: 25/100 (low) - allow' }