Apisecure, 14 & 15 March 2023 - Meenakshi Kaushik, Leader - Product Management at Cisco.
summary
Compares straightforward JSON regex scanning against a trained sequence model and shows tradeoffs: a hand‑built loop of pattern checks masks values quickly for small rule sets but regex list grows O(n²) and latency explodes; generating training data with Faker and training a two‑layer LSTM RNN yields near‑perfect accuracy on 2,000 synthetic samples after 10 epochs but likely overfits; production requires balanced, diverse samples, retraining or fine‑tuning (Hugging Face or GPT‑based) and model cards plus monitoring for data drift.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:30Detecting sensitive data in APIs
- 6:31Simple regex challenges
- 9:48Regex limitations and complexity
- 11:11ML and NLP for detection
- 14:24Training LSTM model
- 17:38Regex vs ML comparison
- 23:00Model accuracy and false negatives
- 26:55Insider threats via API exfiltration
- 28:20Organizing security teams
- 31:33Q&A
