Neural Networks (NN) provide a powerful method for machine learning training and inference. To effectively train, it is desirable for multiple parties to combine their data – however, doing so conflicts with data privacy. In this work, we provide novel three-party secure computation protocols for various NN building blocks such as matrix multiplication, convolutions, Rectified Linear Units, Maxpool, normalization and so on. This enables us to construct three-party secure protocols for training and inference of several NN architectures such that no single party learns any information about the data. Experimentally, we implement our system over Amazon EC2 servers in different settings. Our work advances the state-of-the-art of secure computation for neural networks in three ways: 1. Scalability: We are the first work to provide neural network training on Convolutional Neural Networks (CNNs) that have an accuracy of > 99% on the MNIST dataset; 2. Performance: For secure inference, our system outperforms prior 2 and 3-server works (SecureML, MiniONN, Chameleon, Gazelle) by 6×-113× (with larger gains obtained in more complex networks). Our total execution times are 2 − 4× faster than even just the online times of these works. For secure training, compared to the only prior work (SecureML) that considered a much smaller fully connected network, our protocols are 79× and 7× faster than their 2 and 3-server protocols. In the WAN setting, these improvements are more dramatic and we obtain an improvement of 553×! 3. Security: Our protocols provide two kinds of security: full security (privacy and correctness) against one semi-honest corruption and the notion of privacy against one malicious corruption [Araki et al. CCS’16]. All prior works only provide semi-honest security and ours is the first system to provide any security against malicious adversaries for the secure computation of complex algorithms such as neural network inference and training. Our gains come from a significant improvement in communication through the elimination of expensive garbled circuits and oblivious transfer protocols.

We consider what constitutes identities in cryptography. Typical examples include your name and your social-security number, or your fingerprint/iris-scan, or your address, or your (non-revoked) publickey coming from some trusted public-key infrastructure. In many situations, however, where you are defines your identity. For example, we know the role of a bank-teller behind a bullet-proof bank window not because she shows us her credentials but by merely knowing her location. In this paper, we initiate the study of cryptographic protocols where the identity (or other credentials and inputs) of a party are derived from its geographic location.We start by considering the central task in this setting, i.e., securely verifying the position of a device. Despite much work in this area, we show that in the Vanilla (or standard) model, the above task (i.e., of secure positioning) is impossible to achieve. In light of the above impossibility result, we then turn to the Bounded Storage Model and formalize and construct information theoretically secure protocols for two fundamental tasks:-Secure Positioning; and -Position Based Key Exchange.We then show that these tasks are in fact universal in this setting -we show how we can use them to realize Secure Multi-Party Computation. Our main contribution in this paper is threefold: to place the problem of secure positioning on a sound theoretical footing; to prove a strong impossibility result that simultaneously shows the insecurity of previous attempts at the problem; and to present positive results by showing that the bounded-storage framework is, in fact, one of the "right" frameworks (there may be others) to study the foundations of position-based cryptography.Full version available on eprint.

In this work, we study position-based cryptography in the quantum setting. The aim is to use the geographical position of a party as its only credential. On the negative side, we show that if adversaries are allowed to share an arbitrarily large entangled quantum state, no secure position-verification is possible at all. To this end, we prove the following very general result. Assume that Alice and Bob hold respectively subsystems A and B of a (possibly) unknown quantum state |ψ ∈ H A ⊗ H B . Their goal is to calculate and share a new state |ϕ = U |ψ , where U is a fixed unitary operation. The question that we ask is how many rounds of mutual communication are needed. It is easy to achieve such a task using two rounds of classical communication, whereas in general, it is impossible with no communication at all.Surprisingly, in case Alice and Bob share enough entanglement to start with and we allow an arbitrarily small failure probability, we show that the same task can be done using a single round of classical communication in which Alice and Bob simultaneously exchange two classical messages. Actually, we prove that a relaxed version of the task can be done with no communication at all, where the task is to compute instead a state |ϕ ′ that coincides with |ϕ = U |ψ up to local operations on A and on B, which are determined by classical information held by Alice and Bob. The one-round scheme for the original task then follows as a simple corollary. We also show that these results generalize to more players. As a consequence, we show a generic attack that breaks any position-verification scheme.On the positive side, we show that if adversaries do not share any entangled quantum state but can compute arbitrary quantum operations, secure position-verification is achievable. Jointly, these results suggest the interesting question whether secure position-verification is possible in case of a bounded amount of entanglement. Our positive result can be interpreted as resolving this question in the simplest case, where the bound is set to zero.In models where secure positioning is achievable, it has a number of interesting applications. For example, it enables secure communication over an insecure channel without having any pre-shared key, with the guarantee that only a party at a specific location can learn the content of the conversation. More generally, we show that in settings where secure position-verification is achievable, other position-based cryptographic schemes are possible as well, such as secure position-based authentication and position-based key agreement.

We present CRYPTFLOW, a first of its kind system that converts TensorFlow inference code into Secure Multi-party Computation (MPC) protocols at the push of a button. To do this, we build three components. Our first component, Athos, is an end-to-end compiler from TensorFlow to a variety of semihonest MPC protocols. The second component, Porthos, is an improved semi-honest 3-party protocol that provides significant speedups for TensorFlow like applications. Finally, to provide malicious secure MPC protocols, our third component, Aramis, is a novel technique that uses hardware with integrity guarantees to convert any semi-honest MPC protocol into an MPC protocol that provides malicious security. The malicious security of the protocols output by Aramis relies on integrity of the hardware and semi-honest security of MPC. Moreover, our system matches the inference accuracy of plaintext TensorFlow.We experimentally demonstrate the power of our system by showing the secure inference of real-world neural networks such as RESNET50 and DENSENET121 over the ImageNet dataset with running times of about 30 seconds for semi-honest security and under two minutes for malicious security. Prior work in the area of secure inference has been limited to semi-honest security of small networks over tiny datasets such as MNIST or CIFAR.

Recently, at Crypto 2008, Boneh, Halevi, Hamburg, and Ostrovsky (BHHO) solved the longstanding open problem of "circular encryption," by presenting a public key encryption scheme and proving that it is semantically secure against key dependent chosen plaintext attack (KDM-CPA security) under standard assumptions (and without resorting to random oracles). However, they left as an open problem that of designing an encryption scheme that simultaneously provides security against both key dependent chosen plaintext and adaptive chosen ciphertext attack (KDM-CCA2 security). In this paper, we solve this problem. First, we show that by applying the Naor-Yung "double encryption" paradigm, one can combine any KDM-CPA secure scheme with any (ordinary) CCA2 secure scheme, along with an appropriate non-interactive zero-knowledge proof, to obtain a KDM-CCA2 secure scheme. Second, we give a concrete instantiation that makes use the above KDM-CPA secure scheme of BHHO, along with a generalization of the Cramer-Shoup CCA2 secure encryption scheme, and recently developed pairing-based NIZK proof systems. This instantiation increases the complexity of the BHHO scheme by just a small constant factor.