International Association
for Cryptologic Research

Recent work of Li and Micciancio (Eurocrypt 2021) has shown that the traditional formulation of indistinguishabiity under chosen plaintext attack (IND-CPA) is not adequate to capture the security of approximate homomorphic encryption against passive adversaries, and identified a stronger IND-CPA^D security definition (IND-CPA with decryption oracles) as the appropriate security target for approximate encryption schemes. We show how to transform any approximate homomorphic encryption scheme achieving the weak IND-CPA security definition, into one which is provably IND-CPA^D secure, offering strong guarantees against realistic passive attacks. The method works by postprocessing the output of the decryption function with a mechanism satisfying an appropriate notion of differentially privacy (DP), adding an amount of noise tailored to the worst-case error growth of the homomorphic computation. We apply these results to the approximate homomorphic encryption scheme of Cheon, Kim, Kim, and Song (CKKS, Asiacrypt 2017), proving that adding Gaussian noise to the output of CKKS decryption suffices to achieve IND-CPA^D security. We precisely quantify how much Gaussian noise must be added by proving nearly matching upper and lower bounds, showing that one cannot hope to significantly reduce the amount of noise added in this post-processing step. Based on our upper and lower bounds, we propose parameters for the counter-measures recently adopted by open-source libraries implementing CKKS. Lastly, we investigate the plausible claim that smaller DP noise parameters might suffice to achieve IND-CPA^D-security for schemes supporting more accurate (dynamic, key dependent) estimates of ciphertext noise during decryption. Perhaps surprisingly, we show that this claim is false, and that DP mechanisms with noise parameters tailored to the error present in a given ciphertext, rather than worst-case error, are vulnerable to IND-CPA^D attacks.

A comparison of two encrypted numbers is an important operation needed in many machine learning applications, for example, decision tree or neural network inference/training. An efficient instantiation of this operation in the context of fully homomorphic encryption (FHE) can be challenging, especially when a relatively high precision is sought. The conventional FHE way of evaluating the comparison operation, which is based on the sign function evaluation using FHEW/TFHE bootstrapping (often referred in literature as programmable bootstrapping), can only support very small precision (practically limited to 4-5 bits or so). For higher precision, the runtime complexity scales linearly with the ciphertext (plaintext) modulus (i.e., exponentially with the modulus bit size). We propose sign function evaluation algorithms that scale logarithmically with the ciphertext (plaintext) modulus, enabling the support of large-precision comparison in practice. Our sign evaluation algorithms are based on an iterative use of homomorphic floor function algorithms, which are also derived in our work. Further, we generalize our procedures for floor function evaluation to arbitrary function evaluation, which can be used to support both small plaintext moduli (directly) and larger plaintext moduli (by using a homomorphic digit decomposition algorithm, also suggested in our work). We implement all these algorithms using the PALISADE lattice cryptography library, introducing several implementation-specific optimizations along the way, and discuss our experimental results.

We present passive attacks against CKKS, the homomorphic encryption scheme for arithmetic on approximate numbers presented at Asiacrypt 2017. The attack is both theoretically efficient (running in expected polynomial time) and very practical, leading to complete key recovery with high probability and very modest running times. We implemented and tested the attack against major open source homomorphic encryption libraries, including HEAAN, SEAL, HElib and PALISADE, and when computing several functions that often arise in applications of the CKKS scheme to machine learning on encrypted data, like mean and variance computations, and approximation of logistic and exponential functions using their Maclaurin series. The attack shows that the traditional formulation of IND-CPA security (or indistinguishability against chosen plaintext attacks) achieved by CKKS does not adequately captures security against passive adversaries when applied to approximate encryption schemes, and that a different, stronger definition is required to evaluate the security of such schemes. We provide a solid theoretical basis for the security evaluation of homomorphic encryption on approximate numbers (against passive attacks) by proposing new definitions, that naturally extend the traditional notion of IND-CPA security to the approximate computation setting. We propose both indistinguishability-based and simulation-based variants, as well as restricted versions of the definitions that limit the order and number of adversarial queries (as may be enforced by some applications). We prove implications and separations among different definitional variants, and discuss possible modifications to CKKS that may serve as a countermeasure to our attacks.

Discrete Gaussian distributions over lattices are central to lattice-based cryptography, and to the computational and mathematical aspects of lattices more broadly. The literature contains a wealth of useful theorems about the behavior of discrete Gaussians under convolutions and related operations. Yet despite their structural similarities, most of these theorems are formally incomparable, and their proofs tend to be monolithic and written nearly “from scratch,” making them unnecessarily hard to verify, understand, and extend. In this work we present a modular framework for analyzing linear operations on discrete Gaussian distributions. The framework abstracts away the particulars of Gaussians, and usually reduces proofs to the choice of appropriate linear transformations and elementary linear algebra. To showcase the approach, we establish several general properties of discrete Gaussians, and show how to obtain all prior convolution theorems (along with some new ones) as straightforward corollaries. As another application, we describe a self-reduction for Learning With Errors (LWE) that uses a fixed number of samples to generate an unlimited number of additional ones (having somewhat larger error). The distinguishing features of our reduction are its simple analysis in our framework, and its exclusive use of discrete Gaussians without any loss in parameters relative to a prior mixed discrete-and-continuous approach. As a contribution of independent interest, for subgaussian random matrices we prove a singular value concentration bound with explicitly stated constants, and we give tighter heuristics for specific distributions that are commonly used for generating lattice trapdoors. These bounds yield improvements in the concrete bit-security estimates for trapdoor lattice cryptosystems.

We present a two-message oblivious transfer protocol achieving statistical sender privacy and computational receiver privacy based on the RLWE assumption for cyclotomic number fields. This work improves upon prior lattice-based statistically sender-private oblivious transfer protocols by reducing the total communication between parties by a factor O(nlogq) for transfer of length O(n) messages. Prior work of Brakerski and Dottling uses transference theorems to show that either a lattice or its dual must have short vectors, the existence of which guarantees lossy encryption for encodings with respect to that lattice, and therefore statistical sender privacy. In the case of ideal lattices from embeddings of cyclotomic integers, the existence of one short vector implies the existence of many, and therefore encryption with respect to either a lattice or its dual is guaranteed to ``lose" more information about the message than can be ensured in the case of general lattices. This additional structure of ideals of cyclotomic integers allows for efficiency improvements beyond those that are typical when moving from the generic to ideal lattice setting, resulting in smaller message sizes for sender and receiver, as well as a protocol that is simpler to describe and analyze.

Many advanced lattice cryptography applications require efficient algorithms for inverting the so-called “gadget” matrices, which are used to formally describe a digit decomposition problem that produces an output with specific (statistical) properties. The common gadget inversion problems are the classical (often binary) digit decomposition, subgaussian decomposition, Learning with Errors (LWE) decoding, and discrete Gaussian sampling. In this work, we build and implement an efficient lattice gadget toolkit that provides a general treatment of gadget matrices and algorithms for their inversion/sampling. The main contribution of our work is a set of new gadget matrices and algorithms for efficient subgaussian sampling that have a number of major theoretical and practical advantages over previously known algorithms. Another contribution deals with efficient algorithms for LWE decoding and discrete Gaussian sampling in the Residue Number System (RNS) representation.We implement the gadget toolkit in PALISADE and evaluate the performance of our algorithms both in terms of runtime and noise growth. We illustrate the improvements due to our algorithms by implementing a concrete complex application, key-policy attribute-based encryption (KP-ABE), which was previously considered impractical for CPU systems (except for a very small number of attributes). Our runtime improvements for the main bottleneck operation based on subgaussian sampling range from 18x (for 2 attributes) to 289x (for 16 attributes; the maximum number supported by a previous implementation). Our results are applicable to a wide range of other advanced applications in lattice cryptography, such as GSW-based homomorphic encryption schemes, leveled fully homomorphic signatures, other forms of ABE, some program obfuscation constructions, and more.

We give an efficient decision procedure that, on input two (acyclic) expressions making arbitrary use of common cryptographic primitives (namely, encryption and pseudorandom generators), determines (in polynomial time) if the two expressions produce computationally indistinguishable distributions for any cryptographic instantiation satisfying the standard security notions of pseudorandomness and indistinguishability under chosen plaintext attack. The procedure works by mapping each expression to a symbolic pattern that captures, in a fully abstract way, the information revealed by the expression to a computationally bounded observer. Our main result shows that if two expressions are mapped to different symbolic patterns, then there are secure pseudorandom generators and encryption schemes for which the two distributions can be distinguished with overwhelming advantage. At the same time if any two (acyclic) expressions are mapped to the same pattern, then the associated distributions are indistinguishable.

We describe a somewhat homomorphic GSW-like encryption scheme, natively encrypting matrices rather than just single elements. This scheme offers much better performance than existing homomorphic encryption schemes for evaluating encrypted (nondeterministic) finite automata (NFAs). Differently from GSW, we do not know how to reduce the security of this scheme from LWE, instead we reduce it from a stronger assumption, that can be thought of as an inhomogeneous variant of the NTRU assumption. This assumption (that we term iNTRU) may be useful and interesting in its own right, and we examine a few of its properties. We also examine methods to encode regular expressions as NFAs, and in particular explore a new optimization problem, motivated by our application to encrypted NFA evaluation. In this problem, we seek to minimize the number of states in an NFA for a given expression, subject to the constraint on the ambiguity of the NFA.

We exemplify and evaluate the use of the equational framework of Micciancio and Tessaro (ITCS 2013) by analyzing a number of concrete Oblivious Transfer protocols: a classic OT transformation to increase the message size, and the recent (so called “simplest”) OT protocol in the random oracle model of Chou and Orlandi (Latincrypt 2015), together with some simple variants. Our analysis uncovers subtle timing bugs or shortcomings in both protocols, or the OT definition typically employed when using them. In the case of the OT length extension transformation, we show that the protocol can be formally proved secure using a revised OT definition and a simple protocol modification. In the case of the “simplest” OT protocol, we show that it cannot be proved secure according to either the original or revised OT definition, in the sense that for any candidate simulator (expressible in the equational framework) there is an environment that distinguishes the real from the ideal system.