Language Support for Secure Software Development with Enclaves

Abstract

Confidential computing is a promising technology for securing code and data-in-use on untrusted host machines, e.g., the cloud. Many hardware vendors offer different implementations of Trusted Execution Environments (TEEs). A TEE is a hardware protected execution environment that allows performing confidential computations over sensitive data on untrusted hosts. Despite the appeal of achieving strong security guarantees against low-level attackers, two challenges hinder the adoption of TEEs. First, developing software in high-level managed languages, e.g., Java or Scala, taking advantage of existing TEEs is complex and error-prone. Second, partitioning an application into components that run inside and outside a TEE may break application-level security policies, resulting in an insecure application when facing a realistic attacker. In this work, we study both these challenges. We present JE, a programming model that seamlessly integrates a TEE, abstracting away low-level programming details such as initialization and loading of data into the TEE. JE only requires developers to add annotations to their programs to enable the execution within the TEE. Drawing on information flow control, we develop a security type system that checks confidentiality and integrity policies against realistic attackers with full control over the code running outside the TEE. We formalize the security type system for the JE core and prove it sound for a semantic characterization of security. We implement JE and the security type system, enable Java programs to run on Intel SGX with strong security guarantees. We evaluate our approach on use cases from the literature, including a battleship game, a secure event processing system, and a popular processing framework for big data, showing that we correctly handle complex cases of partitioning, information flow, declassification, and trust.