Case 85BFC3 · AI security · L3 Applied
Fine-Tuning vs RAG for Sensitive Data
Interview questionA team wants the model to answer from confidential internal documents. From a security view, should they fine-tune or use RAG?
- 01 Answer
- 02 Example
- 03 Trade-offs
What a strong answer covers
Try it out loud first. Then check yourself:
- Fine-tuning bakes data into weights: no per-user access control, and the model may regurgitate it to anyone who asks.
- Deletion is hard: removing one record means retraining, which clashes with erasure duties under DPDP or GDPR.
- RAG keeps data in a store you control: enforce user ACLs at retrieval, update or delete documents, cite sources.
- RAG brings its own risks: indirect injection through documents, poisoned or overshared sources, vector store access.
- Typical call: RAG for permissioned or changing data; fine-tune for style or format, on data everyone may see.
If the interviewer pushes back
- They insist on fine-tuning for accuracy. What would you require before the tuned model ships?
- How would you test whether a fine-tuned model memorised specific confidential records?