Detection System

Detection System

LMIF uses a sophisticated 3-layer detection system to identify potential violations of protected identities.

Overview

The detection system runs whenever:

• A platform checks an identity before avatar creation
• Scheduled scans run on existing content
• A creator boxes their identity (scanning for existing violations)

Three-Layer Defense

┌─────────────────────────────────────────────────────────────┐
│                    DETECTION PIPELINE                        │
├─────────────────────────────────────────────────────────────┤
│                                                              │
│  LAYER 1: Registry Match                                     │
│  ├── Speed: <100ms                                           │
│  ├── Method: Database matching (name + image)                │
│  ├── Output: Confidence score                                │
│  └── Triggers: High confidence → AUTO_FLAG                   │
│                Medium confidence → Layer 2                   │
│                Low confidence → NO_ACTION                    │
│                                                              │
│  LAYER 2: AI Analysis                                        │
│  ├── Speed: 2-5 seconds                                      │
│  ├── Method: Vision + semantic analysis                      │
│  ├── Checks: Name variations, leetspeak, visual similarity   │
│  ├── Parody detection: Identifies satire/parody content      │
│  └── Output: Enhanced confidence + classification            │
│                                                              │
│  LAYER 3: Human Review                                       │
│  ├── Speed: 24-48 hours                                      │
│  ├── Method: Human reviewers in queue                        │
│  ├── Handles: Edge cases, appeals, legal gray areas          │
│  └── Output: Final determination                             │
│                                                              │
└─────────────────────────────────────────────────────────────┘

Layer 1: Registry Match

The fastest layer, optimized for sub-100ms responses.

How It Works

1. Extract name from avatar/content
2. Query the boxed identity registry
3. If name matches, compare reference images
4. Calculate confidence score

Triggers

Confidence	Action
High	AUTO_FLAG - Immediate violation created
Medium	Escalate to Layer 2
Low	NO_ACTION - Allow to proceed

What It Catches

• Exact name matches
• Known name variations (registered with claim)
• Direct image matches

Layer 2: AI Analysis

More sophisticated analysis for ambiguous cases.

How It Works

1. Vision analysis of image content
2. Semantic analysis of name/description
3. Cross-reference with boxed identities
4. Parody/satire detection

What It Catches

• Leetspeak variations (T4ylor Sw1ft)
• Phonetic spellings (Tay Tay)
• Visual similarity without exact name
• Deepfakes and AI-generated likenesses

Parody Detection

Note

LMIF includes parody detection to avoid false positives on legitimate satire. Parody content is flagged separately and may be allowed depending on the creator’s policy.

Factors considered:

• Explicit parody disclaimers
• Obvious satirical context
• Transformative elements
• Commentary/criticism markers

Layer 3: Human Review

For edge cases that require human judgment.

When It’s Used

• Appeals from avatar creators
• Legal gray areas
• High-value/high-stakes decisions
• Disputed parody claims

Review Process

1. Case enters review queue
2. Reviewer examines all evidence
3. Decision made within SLA (24-48 hours)
4. Both parties notified of outcome
5. Appeal path available

Key Rule: Image + Name Together

The system requires BOTH image AND name for blocking:

Scenario	Result
”John Smith” alone	NOT blockable (common name)
“John Smith” + Keanu’s face	BLOCKED
Generic image + “Keanu Reeves”	BLOCKED
Generic image + “John Smith”	NOT blockable

This protects real people who share common names with celebrities.

Detection Actions

Action	Meaning	What Happens
AUTO_FLAG	High confidence violation	Violation created, grace period starts
QUEUE_REVIEW	Needs human review	Added to review queue
NO_ACTION	No match found	Nothing happens
PARODY_DETECTED	Identified as parody	May be allowed depending on policy

Detection API

Platforms can trigger detection scans:

// Check single identity
const result = await lmif.identity.check({
  name: "Celebrity Name",
  imageUrl: "https://example.com/avatar.jpg"
});

// Batch check multiple avatars
const results = await lmif.identity.checkBatch([
  { name: "Name 1", imageUrl: "https://..." },
  { name: "Name 2", imageUrl: "https://..." },
]);

// Trigger scan of existing content
const scan = await lmif.detection.scan({
  platformId: "your_platform_id",
  contentType: "avatars"
});

Response Format

{
  "detected": true,
  "confidence": 0.95,
  "layer": 1,
  "matchedIdentity": {
    "claimId": "claim_xyz789",
    "boxId": "box_abc123",
    "name": "Taylor Swift",
    "policy": "MONETIZE"
  },
  "classification": "EXACT_MATCH",
  "parodyLikelihood": 0.02,
  "action": "AUTO_FLAG"
}

Detection Accuracy

Layer	Speed	Accuracy	Use Case
Layer 1	<100ms	95%+	Real-time checks
Layer 2	2-5s	90%+	Fuzzy matching
Layer 3	24-48h	99%+	Final decisions

Tip

For real-time avatar creation flows, Layer 1 is sufficient. Layer 2 runs automatically when needed. You don’t need to explicitly call different layers.