Computing shortest lattice vectors on special hardware

The shortest vector problem (SVP) in lattices is related to problems in combinatorial optimization, algorithmic number theory, communication theory, and cryptography. In 1996, Ajtai published his breakthrough idea how to create lattice-based one-way functions based on the worst-case hardness of an approximate version of SVP. Worst-case hardness is one of the outstanding properties of all modern lattice-based cryptographic schemes. Furthermore, there are no sub-exponential time algorithms known solving SVP, even on potential, strong quantum computers. These facts distinguish the shortest vector problem as a good basis for modern cryptography. In order to theoretically assess the security of lattice-based cryptosystems, knowledge of the asymptotic runtime of SVP solvers is an important issue. For selection of practical parameters however, the average-case behaviour of these algorithms is at least as important. SVP solvers are applied as subroutine in so-called lattice basis reduction algorithms. These build the cornerstone of the fastest attacks on lattice-based cryptosystems. Therefore, improving SVP algorithms directly affects the fastest practical attacks on lattice-based cryptosystems. Building on existing serial SVP algorithms, this thesis presents multiple approaches towards estimating the practical hardness of the shortest vector problem. We employ various special hardware, ranging from multicore CPUs and graphics cards to “supercomputers” and compute clouds. We develop parallel algorithms and assess their practical running times and scalability. Among others, we present our parallel version of the Extreme Pruning Enumeration algorithm, the currently fastest SVP solver available worldwide. Our implementation set the current records in the SVP challenge, the mostly deployed public SVP solver competition. The influence of our work on the security of lattice-based cryptosystems is twofold. First, we help assessing the strength of worst-case problems that build the theoretical basement of lattice-based cryptography. Second, we show how to improve the fastest practical attacks on these systems in the average case. As further result, we present a variant of the sieving algorithm to solve the shortest vector problem in ideal lattices. Ideal lattices are the most important type of lattices in cryptography. Our algorithm is the first to exploit their special structure, allowing us to find shortest vectors faster than in regular lattices.