Abstract
We give a language-based security treatment of domain-specific languages and compilers for secure multi-party computation, a cryptographic paradigm that. enables collaborative computation over encrypted data. Computations are specified in a core imperative language, as if they were intended to be executed by a trusted-third party, and formally verified against. an information-flow policy modelling (an upper bound to) their leakage. This allows non-experts to assess the impact of performance driven authorized disclosure of intermediate values. Specifications are then compiled to multi-party protocols. We formalize protocol security using (distributed) probabilistic information-flow and prove security-preserving compilation: protocols only leak what. is allowed by the source policy. The proof exploits a natural but previously missing correspondence between simulation-based cryptographic proofs and (composable) probabilistic non-interference. Finally, we extend our framework to justify leakage cancelling, a domain-specific optimization that allows to first write an efficient specification that fails to meet the allowed leakage upper-bound, and then apply a probabilistic preprocessing that brings leakage to the acceptable range.

Abstract

Abstract
The adoption of third-party machine learning (ML) cloud services is highly dependent on the security guarantees and the performance penalty they incur on workloads for model training and inference. This paper explores security/performance trade-offs for the distributed Apache Spark framework and its ML library. Concretely, we build upon a key insight: in specific deployment settings, one can reveal carefully chosen non-sensitive operations (e.g. statistical calculations). This allows us to considerably improve the performance of privacy-preserving solutions without exposing the protocol to pervasive ML attacks. In more detail, we propose Soteria, a system for distributed privacy-preserving ML that leverages Trusted Execution Environments (e.g. Intel SGX) to run computations over sensitive information in isolated containers (enclaves). Unlike previous work, where all ML-related computation is performed at trusted enclaves, we introduce a hybrid scheme, combining computation done inside and outside these enclaves. The experimental evaluation validates that our approach reduces the runtime of ML algorithms by up to 41% when compared to previous related work. Our protocol is accompanied by a security proof and a discussion regarding resilience against a wide spectrum of ML attacks.