Development of a Dual Version of DeepBKZ and Its Application to Solving the LWE Challenge

Lattice basis reduction is a strong tool in cryptanalysis. In 2017, DeepBKZ was proposed as a new variant of BKZ, and it calls LLL with deep insertions (DeepLLL) as a subroutine alternative to LLL. In this paper, we develop a dual version of DeepBKZ (which we call “Dual-DeepBKZ”), to reduce the dual basis of an input basis. For Dual-DeepBKZ, we develop a dual version of DeepLLL, and then combine it with the dual enumeration by Micciancio and Walter. It never computes the dual basis of an input basis, and it is as efficient as the primal DeepBKZ. We also demonstrate that Dual-DeepBKZ solves several instances in the TU Darmstadt LWE challenge. We use Dual-DeepBKZ in the bounded distance decoding (BDD) approach for solving an LWE instance. Our experiments show that Dual-DeepBKZ reduces the cost of Liu-Nguyen’s BDD enumeration more effectively than BKZ. For the LWE instance of \((n, \alpha ) = (40, 0.015)\) (resp., \((n, \alpha ) = (60, 0.005)\)), our results are about 2.2 times (resp., 4.0 times) faster than Xu et al.’s results, for which they used BKZ in the fplll library and the BDD enumeration with extreme pruning while we used linear pruning in our experiments.