Using Large Language Models to Improve Data Loss Prevention in Organizations
summary
Cato Networks deploys a fine‑tuned LLM that classifies sensitive documents by creating embeddings and using triplet loss to push same-category items together; a support set (example: 10 labeled docs) and nearest‑neighbor or weighted‑majority voting labels new files via cosine similarity. Demo shows 485 embedded points visualized with a tensorboard demo, using a radius 0.5 threshold and first ~1,000 tokens per file; policy example blocks uploads to ChatGPT when a resume and PII co‑occur.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 1:03What is DLP
- 3:32Legacy DLP limitations
- 5:23Structured data limits for DLP
- 7:56Understanding Large Language Models
- 9:18Fine-tuning LLMs for text classification
- 10:03Text similarity models and triplet loss
- 12:54Training for document classification
- 15:04Classifying documents with embeddings
- 16:29Demo setup with TensorBoard
- 17:12Visualizing embeddings in TensorBoard
- 18:36Classifying documents with nearest neighbors
- 20:06LLM classification with PII detection
- 21:32Key takeaways
- 22:17Q&A
related talks
47mA Step-by-Step Guide to Securing Large Language Models (LLMs)
46mHow Large Language Models Are Reshaping the Cybersecurity Landscape
23mSecurity and auditing tools in Large Language Models (LLM)
25mFinetuning Large Language Models (LLMs) for Security Log Detections
19mSystem for Scaling Large Language Models
01h 19mBuilding an LLM on your Laptop to...
37mExpanding the Capabilities of LLMs Through Community Collaboration- 37mExpanding the Capabilities of LLMs Through Community Collaboration