PrvTel: Lightweight Models for Private and Accurate Telemetry Data Retention

Network operators rely on telemetry for performance and security analysis, but long-term retention at scale remains difficult due to privacy requirements, resource constraints, and the need for high-fidelity query answers. We present PRV-TEL, a framework for privacy-preserving telemetry retention. Instead of storing raw records, PRVTEL learns a compact generative model using a domain-specialized variational autoencoder. It combines field-aware encodings for NetFlow and cloud telemetry with a correlation-aware objective to preserve cross-field dependencies. To enforce differential privacy (DP) without sacrificing utility, PRVTEL injects structure-aware noise before training, rather than during gradient updates. We prove that PRVTEL satisfies DP based on post-processing theorem. Across six real-world datasets and one synthetic workload, PRVTEL improves query accuracy by up to 60% over prior DP-compliant generative baselines and reduces ownership cost by up to 50× compared to lossless retention.