PrvTel: Lightweight Models for Private and Accurate Telemetry Data Retention
summary
PrvTel builds and ships a compact generative retention system that trains a domain-specialized variational autoencoder on privacy-noised telemetry and keeps only the decoder for long-term storage. It injects structure-aware noise before training using a Bayesian-style factorization and adds a covariance-aware loss to preserve cross-field correlations. Provable differential privacy follows from post-processing. Evaluated on six real datasets and one synthetic workload, it yields up to 60% better query accuracy versus prior DP generative baselines, makes membership inference near-random at ε≈2, and cuts ownership cost 50×.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:23Privacy challenges in telemetry
- 4:13Limitations of retention methods
- 7:45PrvTel design insights
- 10:54System workflow and setup
- 11:55Results on fidelity and tasks
- 13:20Privacy guarantees and resistance
- 14:43Long-term retention cost
- 15:27Key takeaways
related talks
14mToward Provably Private Insights into AI Use
31mIs your model private?
17mPrivate Tuning of LLMs in Practice: From VaultGemma to Custom Fine-Tuning
47mPrivate targeting strategies
30mPrivate Generative Models: Balancing Privacy and Utility in AI
28mPrivacy preserving machine learning
31mScaling privacy embedded distributed ML systems
17mPrivid: Practical, Privacy-Preserving Video Analytics Queries