Convolutional neural networks (CNNs) have emerged as one of the most successful deep learning approaches to image recognition and classification. A recent line of research, which includes zkCNN (ACM CCS ’21), vCNN (Cryptology ePrint Archive), and ZEN (Cryptology ePrint Archive), aims at protecting the privacy of CNN models by developing publicly verifiable proofs of correct classification which do not leak any information about the underlying CNN models themselves. A shared feature of these schemes is that they require the entity constructing the proof to have access to both the model and the input in the clear. In other words, a client holding a potentially sensitive input is required to reveal this input to the entity holding the CNN model, thereby sacrificing his privacy, to be able to obtain a verifiable proof of correct classification. This is in contrast to the security guarantees provided by secure classification considered in privacy-preserving machine learning, which does not require the client to reveal his input to obtain a (non-verifiable) classification. In this paper, we propose a privacy-preserving verifiable CNN scheme that overcomes this limitation of the previous schemes by allowing the client to obtain a classification proof without having to reveal his input. The obtained proof allows the client to selectively reveal properties of the obtained classification and his input, which will be verifiable to any third-party verifier. Our scheme is based on the recent notion of collaborative zk-SNARKs by Ozdemir and Boneh (USENIX ’22). Specifically, we construct a new collaborative zk-SNARK based on Bulletproofs achieving an efficient maliciously secure proof generation protocol. Based on this, we then present an optimized approach to CNN evaluation. Finally, we demonstrate the feasibility of our approach by measuring the performance of our scheme on a CNN for classifying the MNIST dataset.

Continuous authentication has been widely studied to provide high security and usability for mobile devices by continuously monitoring and authenticating users. Recent studies adopt multibiometric fusion for continuous authentication to provide high accuracy even when some of captured biometric data are of a low quality. However, existing continuous fusion approaches are resource-heavy as they rely on all classifiers being activated all the time and may not be suitable for mobile devices.In this paper, we propose a new approach to multibiometric continuous authentication: two-dimensional dynamic fusion. Our key insight is that multibiometric continuous authentication calculates two-dimensional matching scores over classifiers and over time. Based on this, we dynamically select a set of classifiers based on the context in which authentication is taking place, and fuse matching scores by multi-classifier fusion and multi-sample fusion. Through experimental evaluation, we show that our approach provides a better balance between resource usage and accuracy than the existing fusion methods. In particular, we show that our approach provides higher accuracy than the existing methods with the same number of score calculations by adopting multi-sample fusion.

Abstract In this paper, we construct multi-key homomorphic and fully homomorphic encryption (resp. MKHE and MKFHE) schemes with malicious circuit privacy. Our schemes are based on learning with errors (LWE) besides appropriate circular security assumptions. In contrast, the previous maliciously circuit-private MKFHE scheme by Chongchitmate and Ostrovsky (PKC, 2017) is based on the non-standard decisional small polynomial ratio (DSPR) assumption with a super-polynomial modulus, besides ring learning with errors and circular security assumptions. We note that it was shown by Albrecht et al. (CRYPTO, 2016) that there exists a sub-exponential time attack against this type of DSPR assumption. The main building block of our maliciously circuit-private MKFHE scheme is a (plain) MKFHE scheme by Brakerski et al. (TCC, 2017), and the security of our schemes is proven under the hardness of LWE with sub-exponential modulus-to-noise ratio and circular security assumptions related to the Brakerski et al. scheme. Furthermore, based on our MKFHE schemes, we construct four-round multi-party computation (MPC) protocols with circuit privacy against a semi-honest server and malicious clients in the plain model. The protocols are obtained by combining our schemes with a maliciously sender-private oblivious transfer protocol and a circuit garbling scheme, all of which can be instantiated only assuming LWE.

In this paper, we propose memory- and round-efficient protocols for securely evaluating arithmetic primitives. We focus on secure two-party computation over the ring ℤ2k that achieves security against semi-honest adversaries and works in the pre-processing model. Our protocols rely on the unit vectorization technique introduced by Boyle et al. (TCC 2019). The unit vectorization technique provides online-optimal protocols for several fundamental operations in the pre-processing model. However, a relatively large memory cost for correlated randomness is required, which might become an obstacle in a large-scale application. In order to achieve both memory and communication efficiency, we propose a size reduction method that uses unit vectorization only for short-length inputs, and based on this, construct two-round protocols for equality test, detecting the most significant non-zero bit, detecting wrap-around, and less-than comparison. In addition, as applications of these results, we provide practically efficient protocols for integer division, integer square root, integer logarithm, and modular exponentiation.

Parallel computation is an important aspect of multi-party computation, not only in terms of improving efficiency, but also in terms of providing privacy for computation involving conditional branching based on private data. While applying multi-party computation in parallel over several sets of input data is straightforward if the partitioning of the input data into sets is publicly known, the problem becomes much more challenging when this partitioning is private. This setting is relevant to broad class of secure computations, in particular to secure graph and database analysis in which the underlying data (graph or database) is private. In this paper, we consider a general class of functions which can be expressed via the iterative evaluation of a binary associative operation, and propose efficient protocols for evaluating such functions in parallel over privately partitioned input data. Our protocols are optimal in terms of the required number of evaluations of the underlying binary operation (i.e.\ N-1 evaluations for total input size N), while simultaneously achieving a round complexity which is only logarithmic in the total size of the input data (i.e.\ O(łog N)).

Machine Learning (ML) algorithms, especially deep neural networks (DNN), have proven themselves to be extremely useful tools for data analysis, and are increasingly being deployed in systems operating on sensitive data, such as recommendation systems, banking fraud detection, and healthcare systems. This underscores the need for privacy-preserving ML (PPML) systems, and has inspired a line of research into how such systems can be constructed efficiently. However, most prior works on PPML achieve efficiency by requiring advanced ML algorithms to be simplified or substituted with approximated variants that are “MPC-friendly” before multi-party computation (MPC) techniques are applied to obtain a PPML systems. A drawback of this approach is that it requires careful fine-tuning of the combined ML and MPC algorithms, and might lead to less efficient algorithms or inferior quality ML (such as lower prediction accuracy). This is an issue for secure training of DNNs in particular, as this involves several arithmetic algorithms that are thought to be “MPCunfriendly”, namely, integer division, exponentiation, inversion, and square root extraction. In this work, we take a structurally different approach and propose a framework that allows efficient and secure evaluation of full-fledged state-of-the-art ML algorithms via secure multi-party computation. Specifically, we propose secure and efficient protocols for the above seemingly MPC-unfriendly computations (but which are essential to DNN). Our protocols are three-party protocols in the honest-majority setting, and we propose both passively secure and actively secure with abort variants. A notable feature of our protocols is that they simultaneously provide high accuracy and efficiency. This framework enables us to efficiently and securely compute modern ML algorithms such as Adam (Adaptive moment estimation) and the softmax function “as is”, without resorting to approximations. As a result, we obtain secure DNN training that outperforms state-of-the-art threeparty systems; our full training is up to 6.7 times faster than just the online phase of FALCON (Wagh et al. at PETS’21) and up to 4.2 times faster than Dalskov et al. (USENIX’21) on the standard benchmark network for secure training of DNNs. The potential advantage of our approach is even greater when considering more complex realistic networks. To demonstrate this, we perform measurements on real-world DNNs, AlexNet and VGG16, which are large networks containing millions of parameters. The performance of our framework for these networks is up to a factor of 26 ∼ 33 faster for AlexNet and 48 ∼ 51 faster for VGG16 to achieve an accuracy of 60% and 70%, respectively, when compared to FALCON. Even compared to CRYPTGPU (Tan et al. IEEE S&P’21), which is optimized for and runs on powerful GPUs, our framework achieves a factor of 2.1 and 4.1 faster performance, respectively, on these networks.

In this paper we propose efficient two-party protocols for obliviously applying a (possibly random) linear group action to a data set. Our protocols capture various applications such as oblivious shuffles, circular shifts, matrix multiplications, to name just a few. A notable feature enjoyed by our protocols, is that they admit a round-optimal (more precisely, one-round) online computation phase, once an input-independent off-line computation phase has been completed. Our oblivious shuffle is the first to achieve a round-optimal online phase. The most efficient instantiations of our protocols are obtained in the so-called client-aided client-server setting, where the offline phase is run by a semi-honest input party (client) who will then distribute the generated correlated randomness to the computing parties (servers). When comparing the total running time to the previous best two-party oblivious shuffle protocol by Chase et al. (Asiacrypt 2020), our shuffle protocol in this client-aided setting is up to 105 times and 152 times faster, in the LAN and WAN setting, respectively. We additionally show how the Chase et al. protocol (which is a standard two-party protocol) can be modified to leverage the advantages of the client-aided setting, but show that, even doing so, our scheme is still two times faster in the online phase and 1.34 times faster in total on average.

Key encapsulation mechanism (KEM) combiners, recently formalized by Giacon, Heuer, and Poettering (PKC'18), enable hedging against insecure KEMs or weak parameter choices by combining ingredient KEMs into a single KEM that remains secure assuming just one of the underlying ingredient KEMs is secure. This seems particularly relevant when considering quantum-resistant KEMs which are often based on arguably less well-understood hardness assumptions and parameter choices. We propose a new simple KEM combiner based on a one-time secure message authentication code (MAC) and two-time correlated input secure hash. Instantiating the correlated input secure hash with a t-wise independent hash for an appropriate value of t, yields a KEM combiner based on a strictly weaker additional primitive than the standard model construction of Giacon et al. and furthermore removes the need to do n full passes over the encapsulation, where n is the number of ingredient KEMs, which Giacon et al. highlight as a disadvantage of their scheme.

It has recently become apparent that both accidental and maliciously caused randomness failures pose a real and serious threat to the security of cryptographic primitives, and in response, researchers have begone the development of primitives that provide robustness against these. In this paper, however, we focus on standardized, widely available primitives. Specifically, we analyze the RSA-OAEP encryption scheme and RSA-PSS signature schemes, specified in PKCS #1, using the related randomness security notion introduced by Paterson et al. (PKC 2014) and its extension to signature schemes. We show that, under the RSA and Φ-hiding assumptions, RSA-OAEP encryption is related randomness secure for a large class of related randomness functions in the random oracle model, as long as the recipient is honest, and remains secure even when additionally considering malicious recipients, as long as the related randomness functions does not allow the malicious recipients to efficiently compute the randomness used for the honest recipient. We furthermore show that, under the RSA assumption, the RSA-PSS signature scheme is secure for any class of related randomness functions, although with a non-tight security reduction. However, under additional, albeit somewhat restrictive assumptions on the related randomness functions and the adversary, a tight reduction can be recovered. Our results provides some reassurance regarding the use of RSA-OAEP and RSA-PSS in environments where randomness failures might be a concern. Lastly, we note that, unlike RSA-OAEP and RSA-PSS, several other schemes, including RSA-KEM, part of ISO 18033-2, and DHIES, part of IEEE P1363a, are not secure under simple repeated randomness attacks.

Secure two-party comparison plays a crucial role in many privacy-preserving applications, such as privacy-preserving data mining and machine learning. In particular, the available comparison protocols with the appropriate input/output configuration have a significant impact on the performance of these applications. In this paper, we firstly describe a taxonomy of secure two-party comparison protocols which allows us to describe the different configurations used for these protocols in a systematic manner. This taxonomy leads to a total of 216 types of comparison protocols.We then describe conversions among these types. While these conversions are based on known techniques and have explicitly or implicitly been considered previously, we show that a combination of these conversion techniques can be used to convert a perhaps less-known two-party comparison protocol by Nergiz et al. (IEEE SocialCom 2010) into a very efficient protocol in a configuration where the two parties hold shares of the values being compared, and obtain a share of the comparison result. This setting is often used in multi-party computation protocols, and hence in many privacy-preserving applications as well. We furthermore implement the protocol and measure its performance. Our measurement suggests that the protocol outperforms the previously proposed protocols for this input/output configuration, when off-line pre-computation is not permitted.

A variant of public key encryption that promises efficiency gains due to batch processing is multi-recipient public key encryption (MR-PKE). Precisely, in MR-PKE, a dedicated encryption routine takes a vector of messages and a vector of public keys and outputs a vector of ciphertexts, where the latter can be decrypted individually, as in regular PKE. In this paper we revisit the established security notions of MR-PKE and the related primitive MR-KEM. We identify a subtle flaw in a security model by Bellare, Boldyreva, and Staddon, that also appears in later publications by different authors. We further observe that these security models rely on the knowledge-of-secret-key (KOSK) assumption---a requirement that is rarely met in practice. We resolve this situation by proposing strengthened security notions for MR-PKE and MR-KEMs, together with correspondingly secure yet highly efficient schemes. Importantly, our models abstain from restricting the set of considered adversaries in the way prior models did, and in particular do not require the KOSK setting. We prove our constructions secure assuming hardness of the static Diffie-Hellman problem, in the random oracle model.