Adversarial Examples, Defense Methods for Online Fraud Detection
summary
Presents an ensemble adversarial generator that combines gradient methods (fast gradient sign and iterative), iterative optimization (Carlini Wagner), generative networks and universal perturbations to craft realistic poisoning and evasion attacks against a PayPal-scale fraud system (305 million accounts, ~12.4 billion transactions). The ensemble lowers model performance more than any single method, targets three bank-fraud scenarios (inadequate funds, stolen or non-existent accounts), and a deployment with adversarial retrain plus online hard example mining shows measurable performance recovery across three time windows.
AI-written summary · about
chapters (12)
- 0:00Start
- 0:27Why Fraud Is an Adversarial Problem
- 1:23Overview of the Agenda
- 2:21Problem of Fraud Detection
- 4:10Domain-Centric Knowledge and Strategy
- 8:36Adversarial Learning Context
- 20:45Iterative Fast Gradient Sine Method
- 21:56Generative Method
- 22:22Universal Adversarial Perturbations
- 23:04Design of the Adversarial Ensemble Generator Method
- 23:58Comparing Individual Methods versus the Ensemble
- 26:22Conclusions
related talks
19mGenerating Adversarial Examples & Defense Methods For Online Fraud Detection
16mLeveraging ML to Create a Reliable Anti-Fraud System
45mWinning by deception: tactics, techniques, and procedures of adversarial A.I.
21mAdversarial Attacks
16mTowards Adversarial Phishing Detection
38mUsing adversarial samples to robustify your Neural Network Models
36mAdversarial Attacks on Deep Learning
42mDefense Against the Dark Arts and Potions: AI Models Can Be Easily Poisoned