Counterfeit Detection

⏱️ Reading time: 9 min · 📁 Category: Brand Protection & Digital Security

Counterfeit detection is delayed and cannot scale if it relies solely on manual complaint workflows. An effective system requires visual similarity analysis, price anomaly detection and marketplace scanning to work together through an algorithmic monitoring approach.

🧠 Visual Similarity: Hashing / ViT / CLIP

• Perceptual Hashing: Rapid near-duplicate detection with aHash/dHash/pHash, even when colours or scale change.
• Deep Features: Extracting ViT/CLIP-based embeddings to cluster similar product photographs in vector space.
• Fast Search: Millisecond-level nearest-neighbour searches (k-NN) across millions of images using HNSW/FAISS indexes.
• Multiple Angles and Manipulation: Detection robust to changes in angle, cropping, watermarks and lighting; combined logo-matching and shape/pattern signals.

💸 Price Anomalies & Market Dynamics

Images alone are insufficient to distinguish genuine products from counterfeits; price behaviour must also be monitored. Listings suddenly falling X% below the SKU reference price are anomaly signals.

• Statistical Triggers: Outlier detection using Z-scores, IQR and MAD: Median Absolute Deviation.
• Time Series: Moving averages + seasonal adjustment to distinguish campaign and flash-discount errors.
• Source Reliability: A weighted risk score based on seller age, return rate and complaint history.

🛒 Marketplace & Social Platform Scanning

• Multi-Source Collection: Products and media from Trendyol, Hepsiburada, Amazon, AliExpress and Instagram/TikTok stores through official APIs or HTML/JSON feeds.
• Listing Normalisation: Converting title, variant, price, seller ID and images into a unique “listing fingerprint”.
• Automatic Matching: Matching against brand catalogues using image embeddings + textual similarity: BM25 or semantic.
• Alerts & Takedown: Automatically completing platform IP portals, DMCA/complaint forms and cease-and-desist drafts.

🎯 Perspective: “Clone listing detection → UDRP/DMCA evidence package”
B10 Digital combines visual matching evidence — the similarity score between an original image and a counterfeit listing — price anomaly charts, seller history and timestamped screenshots in a single Evidence Package. This package accelerates evidentiary work in platform complaints and UDRP/DMCA processes. 📚 Training with an Ethically Sourced 10,000-Image Dataset

A properly constructed training dataset is key to reducing false positives. An ethically sourced dataset of around 10,000 images is built according to these principles:

• Diversity: Product variants, packaging/label close-ups and different lighting/backgrounds; a mix of stock and user images.
• Augmentation: Cropping, blur, compression, colour shifts and watermark/overlay changes simulate typical counterfeiter manipulation.
• Labelling Quality: Double-blind annotation, adjudication of disagreements and maintaining class balance.
• Privacy & Licensing: Permission/licence checks for third-party images; retention compliant with KVKK/GDPR.

📈 KPIs and Operational Metrics

• Takedown Rate: Percentage of reported listings successfully closed.
• False Positive Cost (FPC): Total operational + reputational cost of incorrectly flagged listings that are actually genuine.
  Example formula: FPC = (Number of false positives × processing time × hourly cost) + seller/partner relationship cost coefficient.
• TTR (Time to Remove): Time from initial detection to platform removal approval: hours/days.
• Precision/Recall: Detection accuracy and coverage, balanced through F1 measurement.

The B10 Approach: We combine visual signals (ViT/CLIP), text and price in an ensemble to produce a Risk Score. Score thresholds differ by platform and country, reducing FPC and increasing takedown rates. Legal/PR integration manages objections through a single workflow.

Request Counterfeit Detection & Takedown Consultancy