Defenses and mitigations for privacy attacks
In the previous sections, we described how key LLM aspects (size, complexity, non-deterministic outputs, and the nuanced interplay between content and instructions) change the nature of adversarial privacy attacks. The mitigations we described in Chapters 8 and 9 apply to LLMs broadly but with a shift in attention.
Most mitigations related to MLOps, anonymization techniques, and differential privacy apply to FMs, fine-tuning, and RAG. With FMs becoming a specialized platform activity, your focus as an application developer will be on the following areas:
- Supplier evaluation in data protection policies and ensuring your data is not used for training.
- Supply-chain assessment of open-access models and their data memorization. Model cards for the models in question will help identify the data used to train and support you in devising red teaming exercises to evaluate training data extraction attacks.
- Risk assessment and legal...
