International Association
for Cryptologic Research

In this work, we introduce the sparse LWE assumption, an assumption that draws inspiration from both Learning with Errors (Regev JACM 10) and Sparse Learning Parity with Noise (Alekhnovich FOCS 02). Exactly like LWE, this assumption posits indistinguishability of $(\mathbf{A}, \mathbf{s}\mathbf{A}+\mathbf{e} \mod p)$ from $(\mathbf{A}, \mathbf{u})$ for a random $\mathbf{u}$ where the secret $\mathbf{s}$, and the error vector $\mathbf{e}$ is generated exactly as in LWE. However, the coefficient matrix $\mathbf{A}$ in sparse LPN is chosen randomly from $\mathbb{Z}^{n\times m}_{p}$ so that each column has Hamming weight exactly $k$ for some small $k$. We study the problem in the regime where $k$ is a constant or polylogarithmic. The primary motivation for proposing this assumption is efficiency. Compared to LWE, the samples can be computed and stored with roughly $O(n/k)$ factor improvement in efficiency. Our results can be summarized as:

Foundations: We show several properties of sparse LWE samples, including: 1) The hardness of LWE/LPN with dimension $k$ implies the hardness of sparse LWE/LPN with sparsity $k$ and arbitrary dimension $n \ge k$. 2) When the number of samples $m=\Omega(n \log p)$, length of the shortest vector of a lattice spanned by rows of a random sparse matrix is large, close to that of a random dense matrix of the same dimension (up to a small constant factor). 3) Trapdoors with small polynomial norm exist for random sparse matrices with dimension $n \times m = O(n \log p)$. 4) Efficient algorithms for sampling such matrices together with trapdoors exist when the dimension is $n \times m = \widetilde{\mathcal{O}}(n^2)$.

Cryptanalysis: We examine the suite of algorithms that have been used to break LWE and sparse LPN. While naively many of the attacks that apply to LWE do not exploit sparsity, we consider natural extensions that make use of sparsity. We propose a model to capture all these attacks. Using this model we suggest heuristics on how to identify concrete parameters. Our initial cryptanalysis suggests that concretely sparse LWE with a modest $k$ and slightly bigger dimension than LWE will satisfy similar level of security as LWE with similar parameters.

Applications: We show that the hardness of sparse LWE implies very efficient homomorphic encryption schemes for low degree computations. We obtain the first secret key Linearly Homomorphic Encryption (LHE) schemes with slightly super-constant, or even constant, overhead, which further has applications to private information retrieval, private set intersection, etc. We also obtain secret key homomorphic encryption for arbitrary constant-degree polynomials with slightly super-constant, or constant, overhead.

We stress that our results are preliminary. However, our results make a strong case for further investigation of sparse LWE.

We show that the outsourcing algorithm for the case of linear constraints [IEEE Trans. Cloud Comput., 2023, 11(1), 484-498] cannot keep output privacy, due to the simple translation transformation. We also suggest a remedy method by adopting a hybrid transformation which combines the usual translation transformation and resizing transformation so as to protect the output privacy.

This paper presents a Generalized BFV (GBFV) fully homomorphic encryption scheme that encrypts plaintext spaces of the form $\mathbb{Z}[x]/(\Phi_m(x), t(x))$ with $\Phi_m(x)$ the $m$-th cyclotomic polynomial and $t(x)$ an arbitrary polynomial. GBFV encompasses both BFV where $t(x) = p$ is a constant, and the CLPX scheme (CT-RSA 2018) where $m = 2^k$ and $t(x) = x-b$ is a linear polynomial. The latter can encrypt a single huge integer modulo $\Phi_m(b)$, has much lower noise growth than BFV (linear in $m$ instead of exponential), but cannot be bootstrapped.

We show that by a clever choice of $m$ and higher degree polynomial $t(x)$, our scheme combines the SIMD capabilities of BFV with the low noise growth of CLPX, whilst still being efficiently bootstrappable. Moreover, we present parameter families that natively accommodate packed plaintext spaces defined by a large cyclotomic prime, such as the Fermat prime $\Phi_2(2^{16}) = 2^{16} + 1$ and the Goldilocks prime $\Phi_6(2^{32}) = 2^{64} - 2^{32} + 1$. These primes are often used in homomorphic encryption applications and zero-knowledge proof systems.

Due to the lower noise growth, e.g. for the Goldilocks prime, GBFV can evaluate circuits whose multiplicative depth is more than $5$ times larger than native BFV. As a result, we can evaluate either larger circuits or work with much smaller ring dimensions. In particular, we can natively bootstrap GBFV at 128-bit security for a large prime, already at ring dimension $2^{14}$, which was impossible before. We implemented the GBFV scheme on top of the SEAL library and achieve a latency of only $5$ seconds to bootstrap a ciphertext encrypting $4096$ elements modulo $2^{16}+1$.

We introduce WHIR, a new IOP of proximity that offers small query complexity and exceptionally fast verification time. The WHIR verifier typically runs in a few hundred microseconds, whereas other verifiers in the literature require several milliseconds (if not much more). This significantly improves the state of the art in verifier time for hash-based SNARGs (and beyond). Crucially, WHIR is an IOP of proximity for constrained Reed–Solomon codes, which can express a rich class of queries to multilinear polynomials and to univariate polynomials. In particular, WHIR serves as a direct replacement for protocols like FRI, STIR, BaseFold, and others. Leveraging the rich queries supported by WHIR and a new compiler for multilinear polynomial IOPs, we obtain a highly efficient SNARG for generalized R1CS. As a comparison point, our techniques also yield state-of-the-art constructions of hash-based (non-interactive) polynomial commitment schemes for both univariate and multivariate polynomials (since sumcheck queries naturally express polynomial evaluations). For example, if we use WHIR to construct a polynomial commitment scheme for degree 222, with 100 bits of security, then the time to commit and open is 1.2 seconds, the sender communicates 63 KiB to the receiver, and the opening verification time is 360 microseconds.

We construct a quantum money/quantum lightning scheme from class group actions on elliptic curves over $F_{p}$. Our scheme, which is based on the invariant money construction of Liu-Montgomery-Zhandry (Eurocrypt '23), is simple to describe. We believe it to be the most instantiable and well-defined quantum money construction known so far. The security of our quantum lightning construction is exactly equivalent to the (conjectured) hardness of constructing two uniform superpositions over elliptic curves in an isogeny class which is acted on simply transitively by an exponentially large ideal class group.

However, we needed to advance the state of the art of isogenies in order to achieve our scheme. In partcular, we show: 1. An efficient (quantum) algorithm for sampling a uniform superposition over a cryptographically large isogeny class. 2. A method for specifying polynomially many generators for the class group so that polynomial-sized products yield an exponential-sized subset of class group, modulo a seemingly very modest assumption.

Achieving these results also requires us to advance the state of the art of the (pure) mathematics of elliptic curves, and we are optimistic that the mathematical tools we developed in this paper can be used to advance isogeny-based cryptography in other ways.

We develop and implement AlgoROM, a tool to systematically analyze the security of a wide class of symmetric primitives in idealized models of computation. The schemes that we consider are those that can be expressed over an alphabet consisting of XOR and function symbols for hash functions, permutations, or block ciphers.

We implement our framework in OCaml and apply it to a number of prominent constructions, which include the Luby–Rackoff (LR), key-alternating Feistel (KAF), and iterated Even–Mansour (EM) ciphers, as well as substitution-permutation networks (SPN). The security models we consider are (S)PRP, and strengthenings thereof under related-key (RK), key-dependent message (KD), and more generally key-correlated (KC) attacks.

Using AlgoROM, we are able to reconfirm a number of classical and previously established security theorems, and in one case we identify a gap in a proof from the literature (Connolly et al., ToSC'19). However, most results that we prove with AlgoROM are new. In particular, we obtain new positive results for LR, KAF, EM, and SPN in the above models. Our results better reflect the configurations actually implemented in practice, as they use a single idealized primitive. In contrast to many existing tools, our automated proofs do not operate in symbolic models, but rather in the standard probabilistic model for cryptography.

Oblivious Transfer (OT) is one of the fundamental building blocks in cryptography that enables various privacy-preserving applications. Constructing efficient OT schemes has been an active research area. This paper presents three efficient two-round pairing-free k-out-of-N oblivious transfer protocols with standard security. Our constructions follow the minimal communication pattern: the receiver sends k messages to the sender, who responds with n+k messages, achieving the lowest data transmission among pairing-free k-out-of-n OT schemes. Furthermore, our protocols support adaptivity and also, enable the sender to encrypt the n messages offline, independent of the receiver's variables, offering significant performance advantages in one-sender-multiple-receiver scenarios. We provide security proofs under the Computational Diffie-Hellman (CDH) and RSA assumptions, without relying on the Random Oracle Model. Our protocols combine minimal communication rounds, adaptivity, offline encryption capability, and provable security, making them well-suited for privacy-preserving applications requiring efficient oblivious transfer. Furthermore, the first two proposed schemes require only one operation, making them ideal for resource-constrained devices.

We study differential additions formulas on Kummer lines that factorize through a degree $2$ isogeny $\phi$. We call the resulting formulas half differential additions: from the knowledge of $\phi(P), \phi(Q)$ and $P-Q$, the half differential addition allows to recover $P+Q$. We explain how Mumford's theta group theory allows, in any model of Kummer lines, to find a basis of the half differential relations. This involves studying the dimension $2$ isogeny $(P, Q) \mapsto (P+Q, P-Q)$.

We then use the half differential addition formulas to build a new type of Montgomery ladder, called the half-ladder, using a time-memory trade-off. On a Montgomery curve with full rational $2$-torsion, our half ladder first build a succession of isogeny images $P_i=\phi_i(P_{i-1})$, which only depends on the base point $P$ and not the scalar $n$, for a pre-computation cost of $2S+1m_0$ by bit. Then we use half doublings and half differential additions to compute any scalar multiplication $n \cdot P$, for a cost of $4M+2S+1m_0$ by bit. The total cost is then $4M+4S+2m_0$, even when the base point $P$ is not normalized. By contrast, the usual Montgomery ladder costs $4M+4S+1m+1m_0$ by bit, for a normalized point.

In the appendix, we extend our approach to higher dimensional ladders in theta coordinates.

We present $\mathsf{Protoss}$, a new balanced PAKE protocol with optimal communication efficiency. Messages are only 160 bits long and the computational complexity is lower than all previous approaches. Our protocol is proven secure in the random oracle model and features a security proof in a strong security model with multiple parties and multiple sessions, while allowing for generous attack queries including multiple $\mathsf{Test}$-queries. Moreover, the proof is in the practically relevant single-bit model (that is harder to achieve than the multiple-bit model) and tightly reduces to the Strong Square Diffie-Hellman assumption (SSQRDH). This allows for very efficient, theoretically-sound instantiations and tight compositions with symmetric primitives.

Deep neural networks (DNNs) are valuable assets, yet their public accessibility raises security concerns about parameter extraction by malicious actors. Recent work by Carlini et al. (Crypto’20) and Canales- Martínez et al. (Eurocrypt’24) has drawn parallels between this issue and block cipher key extraction via chosen plaintext attacks. Leveraging differential cryptanalysis, they demonstrated that all the weights and biases of black-box ReLU-based DNNs could be inferred using a polynomial number of queries and computational time. However, their attacks relied on the availability of the exact numeric value of output logits, which allowed the calculation of their derivatives. To overcome this limitation, Chen et al. (Asiacrypt’24) tackled the more realistic hard-label scenario, where only the final classification label (e.g., "dog" or "car") is accessible to the attacker. They proposed an extraction method requiring a polynomial number of queries but an exponential execution time. In addition, their approach was applicable only to a restricted set of architectures, could deal only with binary classifiers, and was demonstrated only on tiny neural networks with up to four neurons split among up to two hidden layers. This paper introduces new techniques that, for the first time, achieve cryptanalytic extraction of DNN parameters in the most challenging hard-label setting, using both a polynomial number of queries and polynomial time. We validate our approach by extracting nearly one million parameters from a DNN trained on the CIFAR-10 dataset, comprising 832 neurons in four hidden layers. Our results reveal the surprising fact that all the weights of a ReLU-based DNN can be efficiently determined by analyzing only the geometric shape of its decision boundaries.

We revisit the problem of Authorized Private Set Intersection (APSI), which allows mutually untrusting parties to authorize their items using a trusted third-party judge before privately computing the intersection. We also initiate the study of Partial-APSI, a novel privacy-preserving generalization of APSI in which the client only reveals a subset of their items to a third-party semi-honest judge for authorization. Partial-APSI allows for partial verification of the set, preserving the privacy of the party whose items are being verified. Both APSI and Partial-APSI have a number of applications, including genome matching, ad conversion, and compliance with privacy policies such as the GDPR. We present two protocols based on bilinear pairings with linear communication. The first realizes the APSI functionality, is secure against a malicious client, and requires only one round of communication during the online phase. Our second protocol realizes the Partial-APSI functionality and is secure against a client that may maliciously inject elements into its input set, but who follows the protocol semi-honestly otherwise. We formally prove correctness and security of these protocols and provide an experimental evaluation to demonstrate their practicality. Our protocols can be efficiently run on commodity hardware. We also show that our protocols are massively parallelizable by running our experiments on a compute grid across 50 cores.

In quantum cryptography, there could be a new world, Microcrypt, where cryptography is possible but one-way functions (OWFs) do not exist. Although many fundamental primitives and useful applications have been found in Microcrypt, they lack ``OWFs-free'' concrete hardness assumptions on which they are based. In classical cryptography, many hardness assumptions on concrete mathematical problems have been introduced, such as the discrete logarithm (DL) problems or the decisional Diffie-Hellman (DDH) problems on concrete group structures related to finite fields or elliptic curves. They are then abstracted to generic hardness assumptions such as the DL and DDH assumptions over group actions. Finally, based on these generic assumptions, primitives and applications are constructed. The goal of the present paper is to introduce several abstracted generic hardness assumptions in Microcrypt, which could connect the concrete mathematical hardness assumptions with applications. Our assumptions are based on a quantum analogue of group actions. A group action is a tuple $(G,S,\star)$ of a group $G$, a set $S$, and an operation $\star:G\times S\to S$. We introduce a quantum analogue of group actions, which we call quantum group actions (QGAs), where $G$ is a set of unitary operators, $S$ is a set of states, and $\star$ is the application of a unitary on a state. By endowing QGAs with some reasonable hardness assumptions, we introduce a natural quantum analogue of the decisional Diffie-Hellman (DDH) assumption and pseudorandom group actions. Based on these assumptions, we construct classical-query pseudorandom function-like state generators (PRFSGs). PRFSGs are a quantum analogue of pseudorandom functions (PRFs), and have many applications such as IND-CPA SKE, EUF-CMA MAC, and private-key quantum money schemes. Because classical group actions are instantiated with many concrete mathematical hardness assumptions, our QGAs could also have some concrete (even OWFs-free) instantiations.

We examine the problem of finding small solutions to systems of modular multivariate polynomials. While the case of univariate polynomials has been well understood since Coppersmith's original 1996 work, multivariate systems typically rely on carefully crafted shift polynomials and significant manual analysis of the resulting Coppersmith lattice. In this work, we develop several algorithms that make such hand-crafted strategies obsolete. We first use the theory of Gröbner bases to develop an algorithm that provably computes an optimal set of shift polynomials, and we use lattice theory to construct a lattice which provably contains all desired short vectors. While this strategy is usable in practice, the resulting lattice often has large rank. Next, we propose a heuristic strategy based on graph optimization algorithms that quickly identifies low-rank alternatives. Third, we develop a strategy which symbolically precomputes shift polynomials, and we use the theory of polytopes to polynomially bound the running time. Like Meers and Nowakowski's automated method, our precomputation strategy enables heuristically and automatically determining asymptotic bounds. We evaluate our new strategies on over a dozen previously studied Coppersmith problems. In all cases, our unified approach achieves the same recovery bounds in practice as prior work, even improving the practical bounds for four of the problems. In four problems, we find smaller and more efficient lattice constructions, and in two problems, we improve the existing asymptotic bounds. While our strategies are still heuristic, they are simple to describe, implement, and execute, and we hope that they drastically simplify the application of Coppersmith's method to systems of multivariate polynomials.

Value Added Tax (VAT) is a cornerstone of government rev- enue systems worldwide, yet its self-reported nature has historically been vulnerable to fraud. While transaction-level reporting requirements may tackle fraud, they raise concerns regarding data security and overreliance on tax authorities as fully trusted intermediaries. To address these issues, we propose Verifiable VAT, a protocol that enables confidential and verifiable VAT reporting. Our system allows companies to confidentially report VAT as a homomorphic commitment in a centrally managed permissioned ledger, using zero-knowledge proofs to provide integrity guarantees. We demonstrate that the scheme strictly limits the amount of fraud possible due to misreporting. Additionally, we introduce a scheme so companies can (dis)prove exchange of VAT with fraudulent companies. The proposed protocol is flexible with regards to real-world jurisdictions’ requirements, and underscores the potential of cryptographic methods to enhance the integrity and confidentiality of tax systems.

We propose a new cryptographic primitive called ``selective batched identity-based encryption'' (Selective Batched IBE) and its thresholdized version. The new primitive allows encrypting messages with specific identities and batch labels, where the latter can represent, for example, a block number on a blockchain. Given an arbitrary subset of identities for a particular batch, our primitive enables efficient issuance of a single decryption key that can be used to decrypt all ciphertexts having identities that are included in the subset while preserving the privacy of all ciphertexts having identities that are excluded from the subset. At the heart of our construction is a new technique that enables public aggregation (i.e. without knowledge of any secrets) of any subset of identities, into a succinct digest. This digest is used to derive, via a master secret key, a single succinct decryption key for all the identities that were digested in this batch. In a threshold system, where the master key is distributed as secret shares among multiple authorities, our method significantly reduces the communication (and in some cases, computation) overhead for the authorities. It achieves this by making their costs for key issuance independent of the batch size.

We present a concrete instantiation of a Selective Batched IBE scheme based on the KZG polynomial commitment scheme by Kate et al. (Asiacrypt'10) and a modified form of the BLS signature scheme by Boneh et al. (Asiacrypt'01). The construction is proven secure in the generic group model (GGM).

In a blockchain setting, the new construction can be used for achieving mempool privacy by encrypting transactions to a block, opening only the transactions included in a given block and hiding the transactions that are not included in it. With the thresholdized version, multiple authorities (validators) can collaboratively manage the decryption process. Other possible applications include scalable support via blockchain for fairness of dishonest majority MPC, and conditional batched threshold decryption that can be used for implementing secure Dutch auctions and privacy preserving options trading.

In this paper, we construct the first asymptotically efficient two-round $n$-out-of-$n$ and multi-signature schemes from lattices in the quantum random oracle model (QROM), using the Fiat-Shamir with Aborts (FSwA) paradigm. Our protocols can be viewed as the QROM~variants of the two-round protocols by Damgård et al. (JoC 2022). A notable feature of our protocol, compared to other counterparts in the classical random oracle model, is that each party performs an independent abort and still outputs a signature in exactly two rounds, making our schemes significantly more scalable.

From a technical perspective, the simulation of QROM~and the efficient reduction from breaking underlying assumption to forging signatures are the essential challenges to achieving efficient QROM security for the previously related works. In order to conquer the former one we adopt the quantum-accessible pseudorandom function (QPRF) to simulate QROM. Particularly, we show that there exist a QPRF~which can be programmed and inverted, even against a quantum adversary. For the latter challenge, we tweak and apply the online extractability by Unruh (Eurocrypt 2015).

Artificial Intelligence (AI) has steadily improved across a wide range of tasks, and a significant breakthrough towards general intelligence was achieved with the rise of generative deep models, which have garnered worldwide attention. However, the development and deployment of AI are almost entirely controlled by a few powerful organizations and individuals who are racing to create Artificial General Intelligence (AGI). These centralized entities make decisions with little public oversight, shaping the future of humanity, often with unforeseen consequences. In this paper, we propose OML, which stands for Open, Monetizable, and Loyal AI, an approach designed to democratize AI development and shift control away from these monopolistic actors. OML is realized through an interdisciplinary framework spanning AI, blockchain, and cryptography. We present several ideas for constructing OML systems using technologies such as Trusted Execution Environments (TEE), traditional cryptographic primitives like fully homomorphic encryption and functional encryption, obfuscation, and AI-native solutions rooted in the sample complexity and intrinsic hardness of AI tasks. A key innovation of our work is the introduction of a new scientific field: AI-native cryptography, which leverages cryptographic primitives tailored to AI applications. Unlike conventional cryptography, which focuses on discrete data and binary security guarantees, AI-native cryptography exploits the continuous nature of AI data representations and their low-dimensional manifolds, focusing on improving approximate performance. One core idea is to transform AI attack methods, such as data poisoning, into security tools. This novel approach serves as a foundation for OML 1.0, an implemented system that demonstrates the practical viability of AI-native cryptographic techniques. At the heart of OML 1.0 is the concept of model fingerprinting, a novel AI-native cryptographic primitive that helps protect the integrity and ownership of AI models. The spirit of OML is to establish a decentralized, open, and transparent platform for AI development, enabling the community to contribute, monetize, and take ownership of AI models. By decentralizing control and ensuring transparency through blockchain technology, OML prevents the concentration of power and provides accountability in AI development that has not been possible before. To the best of our knowledge, this paper is the first to: • Identify the monopolization and lack of transparency challenges in AI deployment today and formulate the challenge as OML (Open, Monetizable, Loyal). • Provide an interdisciplinary approach to solving the OML challenge, incorporating ideas from AI, blockchain, and cryptography. • Introduce and formally define the new scientific field of AI-native cryptography. • Develop novel AI-native cryptographic primitives and implement them in OML 1.0, analyzing their security and effectiveness. • Leverage blockchain technology to host OML solutions, ensuring transparency, decentralization, and alignment with the goals of democratized AI development. Through OML, we aim to provide a decentralized framework for AI development that prioritizes open collaboration, ownership rights, and transparency, ultimately fostering a more inclusive AI ecosystem.

As an emerging primitive, Registered Functional Encryption (RFE) eliminates the key-escrow issue that threatens numerous works for functional encryption, by replacing the trusted authority with a transparent key curator and allowing each user to sample their decryption keys locally. In this work, we present a new black-box approach to construct RFE for all polynomial-sized circuits. It considers adaptive simulation-based security in the bounded collusion model (Gorbunov et al. - CRYPTO'12), where the security can be ensured only if there are no more than Q >= 1 corrupted users and $Q$ is fixed at the setup phase. Unlike earlier works, we do not employ unpractical Indistinguishability Obfuscation (iO). Conversely, it can be extended to support unbounded users, which is previously only known from iO.

Technically, our general compiler exploits garbled circuits and a novel variant of slotted Registered Broadcast Encryption (RBE), namely global slotted RBE. This primitive is similar to slotted RBE, but needs optimally compact public parameters and ciphertext, so as to satisfy the efficiency requirement of the resulting RFE. Then we present two concrete global slotted RBE from pairings and lattices, respectively. With proposed compiler, we hence obtain two bounded collusion-resistant RFE schemes. Here, the first scheme relies on k-Lin assumption, while the second one supports unbounded users under LWE and evasive LWE assumptions.

In this writeup we discuss the soundness of the Basefold multilinear polynomial commitment scheme [Zeilberger, Chen, Fisch 23] applied to Reed-Solomon codes, and run with proximity parameters up to the Johnson list decoding bound. Our security analysis relies on a generalization of the celebrated correlated agreement theorem from [Ben-Sasson, et al., 20] to linear subcodes of Reed-Solomon codes, which turns out a by-product of the Guruswami-Sudan list decoder analysis. We further highlight a non-linear variant of the subcode correlated agreement theorem, which is flexible enough to apply to Basefold-like protocols such as recent optimizations of FRI-Binius [Diamond, Posen 24], and which we believe sufficient for proving the security of a recent multilinear version of STIR [Arnon, Chiesa, Fenzi, Yogev 24] in the list-decoding regime

Recently, deep learning-based side-channel analysis (DLSCA) has emerged as a serious threat against cryptographic implementations. These methods can efficiently break implementations protected with various countermeasures while needing limited manual intervention. To effectively protect implementation, it is therefore crucial to be able to interpret \textbf{how} these models are defeating countermeasures. Several works have attempted to gain a better understanding of the mechanics of these models. However, a fine-grained description remains elusive. To help tackle this challenge, we propose using Kolmogorov-Arnold Networks (KANs). These neural networks were recently introduced and showed competitive performance to multilayer perceptrons (MLPs) on small-scale tasks while being easier to interpret. In this work, we show that KANs are well suited to SCA, performing similarly to MLPs across both simulated and real-world traces. Furthermore, we find specific strategies that the trained models learn for combining mask shares and are able to measure what points in the trace are relevant.