TERMinator Suite: Benchmarking Privacy-Preserving Architectures

Security and privacy are fundamental objectives characterizing contemporary cloud computing. Despite the wide adoption of encryption for protecting data in transit and at rest, data in use remains unencrypted inside cloud processors and memories, as computation is not applicable on encrypted values. This limitation introduces security risks, as unencrypted values can be leaked through side-channels or hardware Trojans. To address this problem, encrypted architectures have recently been proposed, which leverage homomorphic encryption to natively process encrypted data using datapaths of thousands of bits. In this case, additional security protections are traded for higher performance penalties, which drives the need for more efficient architectures. In this work, we develop benchmarks specifically tailored to homomorphic computers, to enable comparisons across different architectures. Our benchmark suite, dubbed TERMinator, is unique as it avoids 'termination problems' that prohibit making control-flow decisions and evaluating early termination conditions based on encrypted data, as these can leak information. Contrary to generic suites that ignore the fundamental challenges of encrypted computation, our algorithms are tailored to the security primitives of the target encrypted architecture, such as the existence of branching oracles. In our experiments, we compiled our benchmarks for the Cryptoleq architecture and evaluated their performance for a range of security parameters.