Detecting Similar ID Documents Using Deep Learning Burkay Gur - PowerPoint PPT Presentation

Detecting Similar ID Documents Using Deep Learning Burkay Gur QCon.ai, Apr 2018

Our mission is to create an 
 open financial system 
 for the world

Risk & Data ● What do we do? - Limit Coinbase’s exposure to risk - Fight Identity Fraud

Attempt 1: Shazam

Attempt 1: Shazam ● Fingerprint for each document ● Perceptual Hashing (256 bit) ● Store hashes in a DB (Hamming distance)

Evaluation of Shazam Pros Cons ● Color differences ● Translations ● Minor cropping ● Large datasets ● Easy to implement ● Domain Specificity

Attempt 2: Vision

Attempt 2: Vision X {

Evaluation of Vision Pros Cons ● Cropping ● Domain Specificity ● Translation ● Iteration Speed ● Infra and Security imgcrypt

New Challenge: Iterate Fast in Highly Secure Environments

Coinbase ML Infrastructure imgcrypt + NostradamusCLI

Takeaways ● Start with naive approach and improve ● Iteration speed is top priority ● Watch out for adversarial attacks Contact: burkay.gur@coinbase.com