(Benford's) Law and Order (Fraud)

on YouTube ↗29m

" summary

Claims that email-username features plus simple models stop most abuse: a custom Sleuth pipeline extracts ~200 features (hex ratio, trigram engram gaps, digit counts, keyboard spread), stores counts in Redis, scores with an online native Bayes classifier, and acts via the Heroku API over a websocket feed. It flags and suspends in about 2 seconds, achieves ~76% detection with 1.5% false positives, and led to ~50,000 suspensions since August.

AI-written summary · about

" chapters (37)

1. 0:00Intro
2. 1:15Assess the situation
3. 2:11Stories are good
4. 3:17Some pretty cool uses.
5. 3:50HEROKU SECURITY OFFICE 650 7TH ST, SAN FRANCISCO THURSDAY, AUG 16, 2013
6. 4:22Run an app for $0
7. 4:400.1% of bad users have a credit card
8. 5:06End-users don't read the TOS
9. 6:41We offer a shell
10. 6:57Really easy to sign up
11. 8:15Static Rules = Hard to maintain
12. 8:27IP Restrictions = Too Hard
13. 8:47Manual Hunting = Unscalable
14. 8:58Manual verification?
15. 10:02Characterset Switches
16. 11:25100% Hex Characters = Hash?
17. 14:31% missing n-grams % bad n-grams Risk Score (0 - 100)
18. 15:24200+ features
19. 15:40Keyboard Spread
20. 16:20My tips for feature extraction?
21. 17:06Shows us where the action is
22. 17:29Operates using (almost) online models
23. 17:54All we need to do is maintain a count of bad vs. good users
24. 18:14We also need to maintain a count of the distinct feature:value pairs for bad vs. good users
25. 20:10Shows a users network
26. 20:46Gives Platform Wide Context
27. 21:20BLOCKING BAD BEHAVIOR
28. 22:39CURRENT ISSUES
29. 23:07Memory usage
30. 24:22Feature selection
31. 24:50Team co-ordination
32. 25:28The occasional mis-classification
33. 25:46Doesn't stop good looking emails
34. 26:13Different Algorithms
35. 26:32CURRENT SUCCESSES
36. 27:50Moar interesting models!
37. 28:12Open Source some things!

" related talks

• 42mBuilding a Robust Fraud Detection SystemGasia Atashian, Gasia Alashian · Data Con LA 2021
• 37mFraud prevention in the age of online privacyAmadou Cisse, Taras Dolinskyi · code.talks 2019
• 13mBehavioral Analytics for Real-Time Fraud PreventionSharath Reddy Polu · Conf42 Cloud Native 2026
• 16mAI-Driven Behavioral Analytics for Real-Time Fraud PreventionSharath Reddy Polu · AI Agents Conference 2026
• 32mDetecting Credential Spearphishing in Enterprise SettingsDavid Wagner, Mobin Javed, Grant Ho, Vern Paxson, Aashish Sharma · 26th USENIX Security Symposium (USENIX Security '17)
• 52mDevil in the HaystackPing Yan · AppSec California 2015
• 21mFinding Spam in Short Text FieldsSayan Dasgupta, Ted Hwa · Spam Fighting @Scale 2015
• 26mIntegrating Identity and FraudAshish Jain · Authenticate 2022