International Association
for Cryptologic Research

In this work, we study GPU implementations of various state-of-the-art sieving algorithms for lattices (Becker-Gama-Joux 2015, Becker-Ducas-Gama-Laarhoven 2016, Herold-Kirshanova 2017) inside the General Sieve Kernel (G6K, Albrecht et al. 2019). In particular, we extensively exploit the recently introduced Tensor Cores -- originally designed for raytracing and machine learning -- and demonstrate their fitness for the cryptanalytic task at hand. We also propose a new dual-hash technique for efficient detection of `lift-worthy' pairs to accelerate a key ingredient of G6K: finding short lifted vectors. We obtain new computational records, reaching dimension 180 for the SVP Darmstadt Challenge improving upon the previous record for dimension 155. This computation ran for 51.6 days on a server with 4 NVIDIA Turing GPUs and 1.5TB of RAM. This corresponds to a gain of about two orders of magnitude over previous records both in terms of wall-clock time and of energy efficiency.

In this talk, I would like to re-popularize two dual ideas that relate Lattices and Factoring. Such a connection may appear surprising at first, but is only one logarithm away: after all, factoring is nothing more than a {\em multiplicative} knapsack problem, i.e. a subset product problem, where the weights are given by the set of small enough primes. The first of the two ideas, we owe to Schnorr (1991) and to Adleman (1995). It consists in finding close or short vectors in a carefully crafted lattice, in the hope that they will provide so-called factoring relations. While this idea does not appear to lead to faster factoring algorithms, it remains fascinating and has in fact lead to other major results. Indeed, the Schnorr-Adleman lattice plays a key role in the proof by Ajtai (1998) of the NP-hardness of the shortest vector problem. The second idea, due to Chor and Rivest (1988) shows a reverse connection: constructing the lattice this time using {\em discrete} logarithms, they instead solve the bounded distance decoding (BDD) problem through easy factoring instances. Revisiting their idea, Pierrot and I (2018) showed that this was a quite close to an optimal construction for solving BDD in polynomial time. It was in fact the best known such construction until some recent work by Peikert and Mook (2020). I wish to conclude with an invitation to explore the cryptographic potential of other lattices than the random q-ary lattices ---the lattices underlying the Learning with Error problem (LWE) and the Short Integer Solution problem (SIS). While SIS and LWE have shown to be very convenient for constructing the most advanced schemes and protocols, I believe that more general lattices have a yet untapped potential for cryptography.

Until recently lattice reduction attacks on NTRU lattices were thought to behave similar as on (ring)-LWE lattices with the same parameters. However several works (Albrecht-Bai-Ducas 2016, Kirchner-Fouque 2017) showed a significant gap for large moduli $q$, the so-called overstretched regime of NTRU. With the NTRU scheme being a finalist to the NIST PQC competition it is important to understand ---both asymptotically and concretely--- where the fatigue point lies exactly, i.e. at which $q$ the overstretched regime begins. Unfortunately the analysis by Kirchner and Fouque is based on an impossibility argument, which only results in an asymptotic upper bound on the fatigue point. It also does not really {\em explain} how lattice reduction actually recovers secret-key information. We propose a new analysis that asymptotically improves on that of Kirchner and Fouque, narrowing down the fatigue point for ternary NTRU from $q \leq n^{2.783+o(1)}$ to $q=n^{2.484+o(1)}$, and finally explaining the mechanism behind this phenomenon. We push this analysis further to a concrete one, settling the fatigue point at $q \approx 0.004 \cdot n^{2.484}$, and allowing precise hardness predictions in the overstretched regime. These predictions are backed by extensive experiments.

The Hidden Subgroup Problem (HSP) aims at capturing all problems that are susceptible to be solvable in quantum polynomial time following the blueprints of Shor's celebrated algorithm. Successful solutions to this problems over various commutative groups allow to efficiently perform number-theoretic tasks such as factoring or finding discrete logarithms. The latest successful generalization (Eisenträger et al. STOC 2014) considers the problem of finding a full-rank lattice as the hidden subgroup of the continuous vector space R^m, even for large dimensions m. It unlocked new cryptanalytic algorithms (Biasse-Song SODA 2016, Cramer et al. EUROCRYPT 2016 and 2017), in particular to find mildly short vectors in ideal lattices. The cryptanalytic relevance of such a problem raises the question of a more refined and quantitative complexity analysis. In the light of the increasing physical difficulty of maintaining a large entanglement of qubits, the degree of concern may be different whether the above algorithm requires only linearly many qubits or a much larger polynomial amount of qubits. This is the question we start addressing with this work. We propose a detailed analysis of (a variation of) the aforementioned HSP algorithm, and conclude on its complexity as a function of all the relevant parameters. Our modular analysis is tailored to support the optimization of future specialization to cases of cryptanalytic interests. We suggest a few ideas in this direction.

Many advanced lattice based cryptosystems require to sample lattice points from Gaussian distributions. One challenge for this task is that all current algorithms resort to floating-point arithmetic (FPA) at some point, which has numerous drawbacks in practice: it requires numerical stability analysis, extra storage for high-precision, lazy/backtracking techniques for efficiency, and may suffer from weak determinism which can completely break certain schemes. In this paper, we give techniques to implement Gaussian sampling over general lattices without using FPA. To this end, we revisit the approach of Peikert, using perturbation sampling. Peikert's approach uses continuous Gaussian sampling and some decomposition $\BSigma = \matA \matA^t$ of the target covariance matrix $\BSigma$. The suggested decomposition, e.g. the Cholesky decomposition, gives rise to a square matrix $\matA$ with real (not integer) entries. Our idea, in a nutshell, is to replace this decomposition by an integral one. While there is in general no integer solution if we restrict $\matA$ to being a square matrix, we show that such a decomposition can be efficiently found by allowing $\matA$ to be wider (say $n \times 9n$). This can be viewed as an extension of Lagrange's four-square theorem to matrices. In addition, we adapt our integral decomposition algorithm to the ring setting: for power-of-2 cyclotomics, we can exploit the tower of rings structure for improved complexity and compactness.

Following the recent line of work on solving the closest vector problem with preprocessing (CVPP) using approximate Voronoi cells, we improve upon previous results in the following ways: We derive sharp asymptotic bounds on the success probability of the randomized slicer, by modelling the behaviour of the algorithm as a random walk on the coset of the lattice of the target vector. We thereby solve the open question left by Doulgerakis–Laarhoven–De Weger [PQCrypto 2019] and Laarhoven [MathCrypt 2019]. We obtain better trade-offs for CVPP and its generalisations (strictly, in certain regimes), both with and without nearest neighbour searching, as a direct result of the above sharp bounds on the success probabilities. We show how to reduce the memory requirement of the slicer, and in particular the corresponding nearest neighbour data structures, using ideas similar to those proposed by Becker–Gama–Joux [Cryptology ePrint Archive, 2015]. Using $$2^{0.185d + o(d)}$$ memory, we can solve a single CVPP instance in $$2^{0.264d + o(d)}$$ time. We further improve on the per-instance time complexities in certain memory regimes, when we are given a sufficiently large batch of CVPP problem instances for the same lattice. Using $$2^{0.208d + o(d)}$$ memory, we can heuristically solve CVPP instances in $$2^{0.234d + o(d)}$$ amortized time, for batches of size at least $$2^{0.058d + o(d)}$$ . Our random walk model for analysing arbitrary-step transition probabilities in complex step-wise algorithms may be of independent interest, both for deriving analytic bounds through convexity arguments, and for computing optimal paths numerically with a shortest path algorithm. As a side result we apply the same random walk model to graph-based nearest neighbour searching, where we improve upon results of Laarhoven [SOCG 2018] by deriving sharp bounds on the success probability of the corresponding greedy search procedure.

Fixing a number field, the space of all ideal lattices, up to isometry, is naturally an Abelian group, called the *Arakelov class group*. This fact, well known to number theorists, has so far not been explicitly used in the literature on lattice-based cryptography. Remarkably, the Arakelov class group is a combination of two groups that have already led to significant cryptanalytic advances: the class group and the unit torus. In the present article, we show that the Arakelov class group has more to offer. We start with the development of a new versatile tool: we prove that, subject to the Riemann Hypothesis for Hecke L-functions, certain random walks on the Arakelov class group have a rapid mixing property. We then exploit this result to relate the average-case and the worst-case of the Shortest Vector Problem in ideal lattices. Our reduction appears particularly sharp: for Hermite-SVP in ideal lattices of certain cyclotomic number fields, it loses no more than a $\tilde O(\sqrt n)$ factor on the Hermite approximation factor. Furthermore, we suggest that this rapid-mixing theorem should find other applications in cryptography and in algorithmic number theory.

We propose a framework for cryptanalysis of lattice-based schemes, when side information --in the form of "hints''-- about the secret and/or error is available. Our framework generalizes the so-called primal lattice reduction attack, and allows the progressive integration of hints before running a final lattice reduction step. Our techniques for integrating hints include sparsifying the lattice, projecting onto and intersecting with hyperplanes, and/or altering the distribution of the secret vector. Our main contribution is to propose a toolbox and a methodology to integrate such hints into lattice reduction attacks and to predict the performance of those lattice attacks with side information. While initially designed for side-channel information, our framework can also be used in other cases: exploiting decryption failures, or simply exploiting constraints imposed by certain schemes (LAC, Round5, NTRU), that were previously not known to (slightly) benefit from lattice attacks. We implement a Sage 9.0 toolkit to actually mount such attacks with hints when computationally feasible, and to predict their performances on larger instances. We provide several end-to-end application examples, such as an improvement of a single trace attack on Frodo by Bos et al (SAC 2018). Contrary to ad-hoc practical attacks exploiting side-channel leakage, our work is a generic way to estimate security loss even given very little side-channel information.

Lattice signature schemes generally require particular care when it comes to preventing secret information from leaking through signature transcript. For example, the Goldreich–Goldwasser–Halevi (GGH) signature scheme and the NTRUSign scheme were completely broken by the parallelepiped-learning attack of Nguyen and Regev (Eurocrypt 2006). Several heuristic countermeasures were also shown vulnerable to similar statistical attacks. At PKC 2008, Plantard, Susilo and Win proposed a new variant of GGH, informally arguing resistance to such attacks. Based on this variant, Plantard, Sipasseuth, Dumondelle and Susilo proposed a concrete signature scheme, called DRS, that is in the round 1 of the NIST post-quantum cryptography project. In this work, we propose yet another statistical attack and demonstrate a weakness of the DRS scheme: one can recover some partial information of the secret key from sufficiently many signatures. One difficulty is that, due to the DRS reduction algorithm, the relation between the statistical leak and the secret seems more intricate. We work around this difficulty by training a statistical model, using a few features that we designed according to a simple heuristic analysis. While we only recover partial secret coefficients, this information is easily exploited by lattice attacks, significantly decreasing their complexity. Concretely, we claim that, provided that $$100\,000$$ 100 000 signatures are available, the secret key may be recovered using BKZ-138 for the first set of DRS parameters submitted to the NIST. This puts the security level of this parameter set below 80-bits (maybe even 70-bits), to be compared to an original claim of 128-bits. Furthermore, we review the DRS v2 scheme that is proposed to resist above statistical attack. For this countermeasure, while one may not recover partial secret coefficients exactly by learning, it seems feasible to gain some information on the secret key. Exploiting this information, we can still effectively reduce the cost of lattice attacks.

We propose the General Sieve Kernel (G6K, pronounced / e.si.ka/), an abstract stateful machine supporting a wide variety of lattice reduction strategies based on sieving algorithms. Using the basic instruction set of this abstract stateful machine, we first give concise formulations of previous sieving strategies from the literature and then propose new ones. We then also give a light variant of BKZ exploiting the features of our abstract stateful machine. This encapsulates several recent suggestions (Ducas at Eurocrypt 2018; Laarhoven and Mariano at PQCrypto 2018) to move beyond treating sieving as a blackbox SVP oracle and to utilise strong lattice reduction as preprocessing for sieving. Furthermore, we propose new tricks to minimise the sieving computation required for a given reduction quality with mechanisms such as recycling vectors between sieves, on-the-fly lifting and flexible insertions akin to Deep LLL and recent variants of Random Sampling Reduction.Moreover, we provide a highly optimised, multi-threaded and tweakable implementation of this machine which we make open-source. We then illustrate the performance of this implementation of our sieving strategies by applying G6K to various lattice challenges. In particular, our approach allows us to solve previously unsolved instances of the Darmstadt SVP (151, 153, 155) and LWE (e.g. (75, 0.005)) challenges. Our solution for the SVP-151 challenge was found 400 times faster than the time reported for the SVP-150 challenge, the previous record. For exact-SVP, we observe a performance crossover between G6K and FPLLL’s state of the art implementation of enumeration at dimension 70.

The hardness of finding short vectors in ideals of cyclotomic number fields (hereafter, Ideal-SVP) can serve as a worst-case assumption for numerous efficient cryptosystems, via the average-case problems Ring-SIS and Ring-LWE. For a while, it could be assumed the Ideal-SVP problem was as hard as the analog problem for general lattices (SVP), even when considering quantum algorithms.But in the last few years, a series of works has lead to a quantum algorithm for Ideal-SVP that outperforms what can be done for general SVP in certain regimes. More precisely, it was demonstrated (under certain hypotheses) that one can find in quantum polynomial time a vector longer by a factor at most $$\alpha = \exp ({\widetilde{O}(n^{1/2})})$$ than the shortest non-zero vector in a cyclotomic ideal lattice, where n is the dimension.In this work, we explore the constants hidden behind this asymptotic claim. While these algorithms have quantum steps, the steps that impact the approximation factor $$\alpha $$ are entirely classical, which allows us to estimate it experimentally using only classical computing. Moreover, we design heuristic improvements for those steps that significantly decrease the hidden factors in practice. Finally, we derive new provable effective lower bounds based on volumetric arguments.This study allows to predict the crossover point with classical lattice reduction algorithms, and thereby determine the relevance of this quantum algorithm in any cryptanalytic context. For example we predict that this quantum algorithm provides shorter vectors than BKZ-300 (roughly the weakest security level of NIST lattice-based candidates) for cyclotomic rings of rank larger than about 24000.

Hash Proof Systems or Smooth Projective Hash Functions (SPHFs) are a form of implicit arguments introduced by Cramer and Shoup at Eurocrypt’02. They have found many applications since then, in particular for authenticated key exchange or honest-verifier zero-knowledge proofs. While they are relatively well understood in group settings, they seem painful to construct directly in the lattice setting.Only one construction of an SPHF over lattices has been proposed in the standard model, by Katz and Vaikuntanathan at Asiacrypt’09. But this construction has an important drawback: it only works for an ad-hoc language of ciphertexts. Concretely, the corresponding decryption procedure needs to be tweaked, now requiring q many trapdoor inversion attempts, where q is the modulus of the underlying Learning With Errors (LWE) problem.Using harmonic analysis, we explain the source of this limitation, and propose a way around it. We show how to construct SPHFs for standard languages of LWE ciphertexts, and explicit our construction over a tag-IND-CCA2 encryption scheme à la Micciancio-Peikert (Eurocrypt’12). We then improve our construction and our analysis in the case where the tag is known in advance or fixed (in the latter case, the scheme is only IND-CPA) with a super-polynomial modulus, to get a stronger type of SPHF, which was never achieved before for any language over lattices.Finally, we conclude with applications of these SPHFs: password-based authenticated key exchange, honest-verifier zero-knowledge proofs, and a relaxed version of witness encryption.

In this paper, we present the lattice-based signature scheme Dilithium, which is a component of the CRYSTALS (Cryptographic Suite for Algebraic Lattices) suite that was submitted to NIST’s call for post-quantum cryptographic standards. The design of the scheme avoids all uses of discrete Gaussian sampling and is easily implementable in constant-time. For the same security levels, our scheme has a public key that is 2.5X smaller than the previously most efficient lattice-based schemes that did not use Gaussians, while having essentially the same signature size. In addition to the new design, we significantly improve the running time of the main component of many lattice-based constructions – the number theoretic transform. Our AVX2-based implementation results in a speed-up of roughly a factor of 2 over the previously best algorithms that appear in the literature. The techniques for obtaining this speed-up also have applications to other lattice-based schemes.

At EUROCRYPT 2013, Garg, Gentry and Halevi proposed a candidate construction (later referred as GGH13) of cryptographic multilinear map (MMap). Despite weaknesses uncovered by Hu and Jia (EUROCRYPT 2016), this candidate is still used for designing obfuscators.The naive version of the GGH13 scheme was deemed susceptible to averaging attacks, i.e., it could suffer from a statistical leak (yet no precise attack was described). A variant was therefore devised, but it remains heuristic. Recently, to obtain MMaps with low noise and modulus, two variants of this countermeasure were developed by Döttling et al. (EPRINT:2016/599).In this work, we propose a systematic study of this statistical leakage for all these GGH13 variants. In particular, we confirm the weakness of the naive version of GGH13. We also show that, among the two variants proposed by Döttling et al., the so-called conservative method is not so effective: it leaks the same value as the unprotected method. Luckily, the leakage is more noisy than in the unprotected method, making the straightforward attack unsuccessful. Additionally, we note that all the other methods also leak values correlated with secrets.As a conclusion, we propose yet another countermeasure, for which this leakage is made unrelated to all secrets. On our way, we also make explicit and tighten the hidden exponents in the size of the parameters, as an effort to assess and improve the efficiency of MMaps.

Lattice signature schemes generally require particular care when it comes to preventing secret information from leaking through signature transcript. For example, the Goldreich-Goldwasser-Halevi (GGH) signature scheme and the NTRUSign scheme were completely broken by the parallelepiped-learning attack of Nguyen and Regev (Eurocrypt 2006). Several heuristic countermeasures were also shown vulnerable to similar statistical attacks.At PKC 2008, Plantard, Susilo and Win proposed a new variant of GGH, informally arguing resistance to such attacks. Based on this variant, Plantard, Sipasseuth, Dumondelle and Susilo proposed a concrete signature scheme, called DRS, that has been accepted in the round 1 of the NIST post-quantum cryptography project.In this work, we propose yet another statistical attack and demonstrate a weakness of the DRS scheme: one can recover some partial information of the secret key from sufficiently many signatures. One difficulty is that, due to the DRS reduction algorithm, the relation between the statistical leak and the secret seems more intricate. We work around this difficulty by training a statistical model, using a few features that we designed according to a simple heuristic analysis.While we only recover partial information on the secret key, this information is easily exploited by lattice attacks, significantly decreasing their complexity. Concretely, we claim that, provided that $$100\,000$$ signatures are available, the secret key may be recovered using BKZ-138 for the first set of DRS parameters submitted to the NIST. This puts the security level of this parameter set below 80-bits (maybe even 70-bits), to be compared to an original claim of 128-bits.