International Association
for Cryptologic Research

Secure 2-party computation (2PC) enables secure inference that offers protection for both proprietary machine learning (ML) models and sensitive inputs to them. However, the existing secure inference solutions suffer from high latency and communication overheads, particularly for transformers. Function secret sharing (FSS) is a recent paradigm for obtaining efficient 2PC protocols with a preprocessing phase. We provide SIGMA, the first end-to-end system for secure transformer inference based on FSS. By constructing new FSS-based protocols for complex machine learning functionalities, such as Softmax and GeLU, and also accelerating their computation on GPUs, SIGMA improves the latency of secure inference of transformers by 11 − 19× over the state-of-the-art that uses preprocessing and GPUs. We present the first secure inference of generative pre-trained transformer (GPT) models. In particular, SIGMA executes GPT-Neo with 1.3 billion parameters in 7.4s and HuggingFace’s GPT2 in 1.6s.

Recently Boyle et al. (TCC 2019) proposed a new approach for secure computation in the {\em preprocessing model} building on {\em function secret sharing} (FSS). This approach can be used to realize any circuit containing gates that admit efficient FSS schemes. In this work, we make the following three technical contributions: {\bf Improved Key Size.} The complexity of the preprocessing phase directly depends on the FSS key size. We improve the size of FSS keys for several existing FSS constructions through two important steps. First, we present a roughly $4\times$ reduction in FSS key size for the Distributed Comparison Function (DCF), i.e. ($f_\alpha(x) = \beta$ for all $x < \alpha$ and $0$, otherwise). Second, prior FSS schemes for many important function classes are obtained via reductions to multiple instances of DCF; for example, 2 instances for interval containment and $2m$ for splines with $m$ pieces. We significantly improve these reductions for public intervals and obtain {\em optimal} FSS schemes, i.e., through a {\em single instance of DCF}, thereby reducing the key sizes by up to $6-22\times$ for commonly used functions in mixed-mode secure computation such as ReLU and sigmoid. {\bf FSS for New Function Families.} We present the first constructions of FSS schemes for arithmetic and logical right shift, as well as for bit-decomposition, where the output bits must be secret shared in a larger ring. These functions are crucial for many applications such as fixed-point arithmetic and machine learning. {\bf FSS for Fixed-Point Arithmetic and Barrier.} One of the important functions in the realization of secure fixed-point arithmetic is that of multiply-then-truncate. While our work shows how to obtain a construction for this function in 2 rounds using sequential calls to FSS schemes for multiply and shift, we demonstrate a barrier towards improving this via FSS beyond what we achieve. Specifically, we show that a 1-round solution would require settling a major open problem in the area of FSS: namely, building an FSS for the class of bit-conjunction functions based on only symmetric-key cryptographic assumptions.

This paper constructs high-rate non-malleable codes in the information-theoretic plain model against tampering functions with bounded locality. We consider $$\delta $$-local tampering functions; namely, each output bit of the tampering function is a function of (at most) $$\delta $$ input bits. This work presents the first explicit and efficient rate-1 non-malleable code for $$\delta $$-local tampering functions, where $$\delta =\xi \lg n$$ and $$\xi <1$$ is any positive constant. As a corollary, we construct the first explicit rate-1 non-malleable code against NC$$^0$$ tampering functions.Before our work, no explicit construction for a constant-rate non-malleable code was known even for the simplest 1-local tampering functions. Ball et al. (EUROCRYPT–2016), and Chattopadhyay and Li (STOC–2017) provided the first explicit non-malleable codes against $$\delta $$-local tampering functions. However, these constructions are rate-0 even when the tampering functions have 1-locality. In the CRS model, Faust et al. (EUROCRYPT–2014) constructed efficient rate-1 non-malleable codes for $$\delta = O(\log n)$$ local tampering functions.Our main result is a general compiler that bootstraps a rate-0 non-malleable code against leaky input and output local tampering functions to construct a rate-1 non-malleable code against $$\xi \lg n$$-local tampering functions, for any positive constant $$\xi < 1$$. Our explicit construction instantiates this compiler using an appropriate encoding by Ball et al. (EUROCRYPT–2016).

Most secure computation protocols can be effortlessly adapted to offload a significant fraction of their computationally and cryptographically expensive components to an offline phase so that the parties can run a fast online phase and perform their intended computation securely. During this offline phase, parties generate private shares of a sample generated from a particular joint distribution, referred to as the correlation. These shares, however, are susceptible to leakage attacks by adversarial parties, which can compromise the security of the secure computation protocol. The objective, therefore, is to preserve the security of the honest party despite the leakage performed by the adversary on her share.Prior solutions, starting with n-bit leaky shares, either used 4 messages or enabled the secure computation of only sub-linear size circuits. Our work presents the first 2-message secure computation protocol for 2-party functionalities that have $$\varTheta (n)$$ circuit-size despite $$\varTheta (n)$$-bits of leakage, a qualitatively optimal result. We compose a suitable 2-message secure computation protocol in parallel with our new 2-message correlation extractor. Correlation extractors, introduced by Ishai, Kushilevitz, Ostrovsky, and Sahai (FOCS–2009) as a natural generalization of privacy amplification and randomness extraction, recover “fresh” correlations from the leaky ones, which are subsequently used by other cryptographic protocols. We construct the first 2-message correlation extractor that produces $$\varTheta (n)$$-bit fresh correlations even after $$\varTheta (n)$$-bit leakage.Our principal technical contribution, which is of potential independent interest, is the construction of a family of multiplication-friendly linear secret sharing schemes that is simultaneously a family of small-bias distributions. We construct this family by randomly “twisting then permuting” appropriate Algebraic Geometry codes over constant-size fields.