Explore perceptual hashing for image spam detection

[vcarl]

Problem

Spammers can evade duplicate message detection by varying attachment filenames, compression settings, or binary data while posting visually identical images across channels.

Current behavior:

• ✅ Catches identical text with varying attachments (content hash is text-only)
• ❌ Misses identical images with varied text (e.g., "check this", "see this", "look here")
• ❌ Misses near-identical images (re-compressed, slightly cropped, filtered)

Proposed Solution

Implement perceptual hashing (pHash/dHash) to detect visually similar images.

How it works

1. Generate perceptual hash for each image attachment (~10-50ms per image)
2. Store hashes in RecentMessage interface
3. Compare incoming image hashes to recent hashes using Hamming distance
4. Flag messages with similar images as potential duplicates

Algorithm Options

dHash (Difference Hash) - Recommended

• Fast, simple, good accuracy for near-duplicates
• Compares adjacent pixel differences
• Node.js: imghash library

pHash (Perceptual Hash)

• More robust, uses DCT (like JPEG compression)
• Slightly slower but better for modified images

PDQ Hash (Facebook)

• Industry standard, open source
• Best accuracy but more complex
• https://github.com/facebook/ThreatExchange/tree/main/pdq

Example Implementation

// Add to RecentMessage interface
export interface RecentMessage {
  messageId: string;
  channelId: string;
  contentHash: string;
  timestamp: number;
  hasLink: boolean;
  imageHashes?: string[]; // NEW: perceptual hashes
}

// In velocityAnalyzer.ts
function countSimilarImages(
  recentMessages: RecentMessage[],
  currentHashes: string[],
  windowMs: number,
  threshold: number = 5, // Hamming distance threshold
): number {
  // Compare hashes, count similar images
}

// New spam signal
if (similarImageCount >= 3) {
  signals.push({
    name: "similar_image_spam",
    score: 5,
    description: `${similarImageCount} visually similar images`,
  });
}

Benefits

• ✅ Catches spammers varying filenames/compression
• ✅ Catches slightly modified images (cropped, filtered, re-encoded)
• ✅ Fast enough for real-time (~10-50ms per image)
• ✅ Low memory footprint (64-bit hash per image)
• ✅ Works with existing velocity detection system

Trade-offs

• ⚠️ Adds latency to message processing (need to download + hash images)
• ⚠️ Won't catch heavily edited images (mirrored, rotated 90°, etc.)
• ⚠️ Requires new dependency (imghash or similar)
• ⚠️ False positives possible (legitimate users posting similar memes)

Recommended Approach

1. Research phase: Test imghash library with sample spam images to validate accuracy
2. Prototype: Implement basic dHash comparison without spam scoring
3. Tune thresholds: Determine optimal Hamming distance threshold (likely 5-10)
4. Add signal: Create similar_image_spam signal with appropriate scoring
5. Monitor: Watch for false positives in production

Libraries to Evaluate

• imghash - Simple wrapper around sharp for perceptual hashing
• sharp - Fast image processing (could implement custom dHash)
• blockhash-js - Alternative perceptual hash implementation
• PDQ hash - Facebook's open source algorithm (more complex but better)

Not Recommended

• AI embeddings (CLIP, etc.) - Too slow and resource-intensive
• PhotoDNA - Requires Microsoft partnership, overkill for spam

Next Steps

• Research and test perceptual hashing libraries
• Collect sample spam images to validate approach
• Prototype implementation
• Determine appropriate scoring and thresholds