Privacy-preserving data analytics on trusted execution environments

Abstract

Modern data-intensive applications often run on platforms that cannot be fully trusted. Sensitive data must therefore be protected during communication, storage, and processing. We study privacy-preserving data analytics in two application areas.

First, we present ASKIndex, a concurrent, disk-aware learned index with a lock-free core. We also present ASKIndexSE, which adds authenticated encrypted persistence, secure client communication, and SGX-enabled execution. ASKIndex achieves up to three times higher throughput than the evaluated disk-aware learned-index baselines. Under the stated threat model, ASKIndexSE protects encrypted-mode query payloads and persistent page contents.

Second, we study federated LSTM forecasting over decentralized smart-energy data. The system protects federated messages with authenticated encryption. It also evaluates local training and server coordination in a Gramine protected-execution simulation. The selected 1-hour configuration reports MAE of 0.4342±0.0016 and RMSE of 0.6856 ± 0.0008 across three random seeds.

Together, the two systems show that privacy mechanisms must match the workload. They also require explicit threat models, careful key management, clear leakage boundaries, and measured performance costs.