PEPR '26 - Toward Provably Private Insights into AI Use @UsenixOrg

Uploaded June 2026 | Updated September 2026, 3 weeks ago

Toward Provably Private Insights into AI Use

Rakshita Tandon, Google

Understanding real-world usage is critical for improving Generative AI, yet traditional analytics often risk exposing sensitive input data. This talk outlines Provably Private Insights (PPI), a novel framework that enables developers to gain deep analytical utility without compromising user privacy. PPI bridges the gap between raw data and actionable insights by integrating Trusted Execution Environments (TEEs) for external transparency and verifiability, "Data Expert" LLMs for interpreting unstructured data within secure enclaves, and Differential Privacy (DP) for mathematically-guaranteed anonymity in aggregation. The talk describes the open-sourced system architecture, and its real-world application in the Recorder app. This framework illustrates the shift beyond classic data analytics toward a "provably private" standard where server-side processing is transparent, verifiable, and restricted to privacy-preserving computations.
Authors: Albert Cheu, Artem Lagzdin, Brett McLarnon, Daniel Ramage, Katharine Daly, Marco Gruteser, Peter Kairouz, Rakshita Tandon, Stanislav Chiknavaryan, Timon Van Overveldt, Zoe Gong

View the full PEPR '26 program at usenix.org/conference/pepr26/program

PEPR 26 - Toward Provably Private Insights into AI Use

PEPR 26 - DPSynth: From Research to Production—Engineering Differentially Private Synthetic...

$NSDI 26 - FRCC: Towards Provably Fair and Robust Congestion Control FRCC: Towards Provably Fair and Robust Congestion Control Anup Agarwal, Carnegie Mellon University; Venkat Arun, University of Texas at Austin; Srinivasan Seshan, Carnegie Mellon University Congestion control algorithms (CCAs) play a critical role in network bandwidth allocation. Recent work (from SIGCOMM 2022) showed that a large class of CCAs, including BBR, Copa, and Reno, starve flows in the presence of network jitter. Starvation occurs because CCAs coordinate fairness by encoding fair rates into congestion signals. For example, Renos throughput scales as 1/√loss rate. Even a small amount of noise in these signals leads to large errors in inferring fair rates. We present FRCC (Fair and Robust Congestion Controller), the first CCA that provably bounds unfairness (avoids starvation) even under network jitter. Our key insight is to encode only the flow count (or equivalently, the fair link fraction) into the congestion signals, and independently estimate the link capacity to calculate the fair rate. In this way, we bound jitters impact on fairness. We implement FRCC in the Linux kernel and evaluate it in a variety of network conditions, including synthetic jitter, heterogeneous RTTs, and multi-bottleneck settings. FRCC closely matches the bounds predicted by our theoretical analysis, and consistently achieves fairness, even when state-of-the-art CCAs exhibit starvation. View the full NSDI 26 program at https://www.usenix.org/conference/nsdi26/technical-sessions$

USENIX Security 25 - ALERT: Machine Learning-Enhanced Risk Estimation for Databases Supporting...

NSDI 26 - cc-pipe: Breaking Systemic Bottlenecks in RPKI Data Supply Chain with Concurrent and...

NSDI 26 - Who Watches the Watchers? On the Reliability of Softwarizing Cloud Application Management

NSDI 26 - SLATE: Service Layer Traffic Engineering

PEPR 26 - Scaling Privacy Threat Modeling: From Architects to Developers

NSDI 26 - FAST: An Efficient Scheduler for All-to-All GPU Communication

NSDI 26 - Express Lane to Efficiency and Reliability: Multi-Dimensional Control in Meta’s Express..

NSDI 26 - CascadeNet: Generating Network Traffic with High-Fidelity Temporal Patterns

PEPR 26 - Privacy Review for Non-Maniacs