Driven by breakthroughs in mobile and IoT devices, on-device computation becomes promising. Meanwhile, there is a growing concern over its security: it faces many threats
in the wild, while not supervised by security experts; the computation is highly likely to touch users’ privacy-sensitive information. Towards trustworthy on-device computation, we present novel system designs focusing on two key applications: stream analytics, and machine learning training and inference.
First, we introduce Streambox-TZ (SBT), a secure stream analytics engine for ARM-based edge platforms. SBT contributes a data plane that isolates only analytics’ data and
computation in a trusted execution environment (TEE). By design, SBT achieves a minimal trusted computing base (TCB) inside TEE, incurring modest security overhead.
Second, we design a minimal GPU software stack (50KB), called GPURip. GPURip allows developers to record GPU computation ahead of time, which will be replayed later
on client devices. In doing so, GPURip excludes the original GPU stack from run time eliminating its wide attack surface and exploitable vulnerabilities.
Finally, we propose CoDry, a novel approach for TEE to record GPU computation remotely. CoDry provides an online GPU recording in a safe and practical way; it hosts GPU stacks in the cloud that collaboratively perform a dryrun with client GPU models. To overcome frequent interactions over a wireless connection, CoDry implements a suite of key optimizations.