Paper 2018/756

Obfuscation Using Tensor Products

Craig Gentry and Charanjit S. Jutla and Daniel Kane

Abstract

We describe obfuscation schemes for matrix-product branching programs that are purely algebraic and employ matrix groups and tensor algebra over a finite field. In contrast to the obfuscation schemes of Garg et al (SICOM 2016) which were based on multilinear maps, these schemes do not use noisy encodings. We prove that there is no efficient attack on our scheme based on re-linearization techniques of Kipnis-Shamir (CRYPTO 99) and its generalization called XL-methodology (Courtois et al, EC2000). We also provide analysis to claim that general Grobner-basis computation attacks will be inefficient. In a generic colored matrix model our construction leads to a virtual-black-box obfuscator for NC$^1$ circuits. We also provide cryptanalysis based on computing tangent spaces of the underlying algebraic sets.

Note: The paper now has Daniel Kane as a co-author. The scheme is slightly revised from the previous scheme and the paper describes an attack on the previous scheme. (Typos fixed from previous version).