Securely and Efficiently Outsourcing Decision Tree Inference

Optimizing Secure Decision Tree Inference Outsourcing

“…We implement our system and make deployment on the Amazon cloud for performance evaluation over various decision trees with realistic sizes. Compared with the state-of-theart prior work [12], the overall online end-to-end inference latency at the cloud servers over realistic WAN environment is up to 8× better. In the meantime, our system offers the provider up to 19× savings in communication and 18× savings in computation.…”

“…Overcoming the privacy hurdles is thus of paramount importance to help the provider and client gain confidence in outsourced decision tree inference services. Towards this challenge, a recent research endeavor has been presented by Zheng et al [12], which represents the state-of-the-art. Their design is based on the lightweight additive secret sharing technique and works under a compatible architecture where two cloud servers from independent cloud providers are employed to jointly conduct the decision tree inference in the ciphertext domain.…”

“…In light of the above observations, In this paper, we present a new highly efficient design for secure decision tree inference outsourcing which significantly improves upon the state-of-the-art. Our design follows the same architecture of [12], and also makes use of additive secret sharing, yet with significant optimizations to achieve largely boosted performance compared to the state-of-the-art work.…”

“…During the procedure, it is required that no information about the feature vector, index value, and selected feature be revealed. With this observation, we propose a new secure feature selection design where the provider only needs to construct and encrypt an indexing vector with size O(J), rather than a matrix of size O(J • I) as proposed in [12].…”

“…Secondly, we note that the linear communication round complexity of the prior work [12] in the secure decision node evaluation phase is due to the secure realization of a ripple carry adder for secret-shared comparison, which faces a delay problem due to sequential procedure of carry computation. Our observation from the field of digital circuit design is that the carry delay problem presented in the ripple carry adder can be solved via the advanced carry look-ahead adder [14].…”