Generating Adversarial Examples & Defense Methods For Online Fraud Detection
summary
Claims that realistic adversarial examples for PayPal-style transaction fraud can be generated by combining ensemble adversarial perturbations (FGSM, L-BFGS, black/white‑box) with domain constraints on categorical and continuous features (e.g., country bin, IP, device) and a second-stage generative cohort optimizer that finds smallest changes that flip decisions. A two-stage defense—adversarial retrain with frozen feedforward then hard‑example backprop—reduces model degradation across five time periods, yields the largest decision-change per perturbation in experiments, and keeps false positives controlled. Results are promising but still evolving.
AI-written summary · about
chapters (7)
- 0:00Introduction to fraud
- 2:38System challenges
- 6:19Adversarial learning
- 9:20Constraints in fraud
- 12:02Attack generation
- 14:22Defensive strategies
- 15:43Q&A and future outlook
related talks
30mAdversarial Examples, Defense Methods for Online Fraud Detection
16mLeveraging ML to Create a Reliable Anti-Fraud System
46mLessons Learned from Evaluating the Robustness of Defenses to Adversarial Examples
18mReal-time Fraud Detection using Deep Learning
38mUsing adversarial samples to robustify your Neural Network Models
20mAttacking and Defending Machine Learning Applications of Public Cloud
21mAdversarial Attacks
29mDeep Learning Applications to Online Payment Fraud Detection