Home

Publications

Research projects

Group meetings

Teaching

Contact

Applied Crypto Group

LACS Crypto Day, June 13th, 2017, room MSA 4.410, Belval campus

• 10:00: Aleksei Udovenko. On division property cryptanalysis.
• 10:30: Dmitry Khovratovich. BIP32-Ed25519: Hierarchical Deterministic Keys over a Non-linear Keyspace.
• 11:00: Benoit Cogliati: Security proofs for symmetric-key constructions
• 11:30: Jun Wang. vdNets: Applying Very Deep Neural Networks to Encrypted Data

  Abstract: Very deep neural networks have recently demonstrated start-of-the-art accuracy on very complex visual and speech recognition tasks. To those problems which involve sensitive data, e.g. medicine and finance, the privacy and security requirements may prevent the use of cloud-based machine learning service. To address this issue, a number of works have explored how to apply machine learning models, including neural networks, to encrypted data. Unfortunately, such works either focus on conventional machine learning models or shallow neural networks. In this paper, we describe the design and implementation of a secure multiparty computation (SMC) system called vdNets which allows performing a learned very deep and large neural network on encrypted data. We first demonstrate the ability of vdNets by applying GoogLeNet convolutional neural network \cite{szegedy2015going} to encrypted data, and analyze its computation and communication cost. GoogLeNet has 22 layers (approximately 100 independent sub-layers) and overall 1.5 billion multiply-adds at inference time. We further discuss the bottleneck of vdNets and introduce different methods to eliminate or alleviate it. Our results provide compelling evidence that an SMC approach to very deep and large neural networks is worth pursuing.

• 14:00: Moon Sung Lee. Attacks on Multilinear Maps
• 14:30: Vincenzo Iovino. The Simplest Oblivious Transfer Protocol.
• 15:00: Alfredo Rial Duran. UC Commitments for Modular Protocol Design and Applications to Revocation and Attribute Tokens

  Abstract: Complex cryptographic protocols are often designed from simple cryptographic primitives, such as signature schemes, encryption schemes, verifiable random functions, and zero-knowledge proofs, by bridging between them with commitments to some of their inputs and outputs. Unfortunately, the known universally composable (UC) functionalities for commitments and the cryptographic primitives mentioned above do not allow such constructions of higher-level protocols as hybrid protocols. Therefore, protocol designers typically resort to primitives with property-based definitions, often resulting in complex monolithic security proofs that are prone to mistakes and hard to verify.

  We address this gap by presenting a UC functionality for non-interactive commitments that enables modular constructions of complex protocols within the UC framework. We also show how the new functionality can be used to construct hybrid protocols that combine different UC functionalities and use commitments to ensure that the same inputs are provided to different functionalities.

  We further provide UC functionalities for attribute tokens and revocation that can be used as building blocks together with our UC commitments. As an example of building a complex system from these new UC building blocks, we provide a construction (a hybrid protocol) of anonymous attribute tokens with revocation. Unlike existing accumulator-based schemes, our scheme allows one to accumulate several revocation lists into a single commitment value and to hide the revocation status of a user from other users and verifiers.

  Co-authors: Jan Camenisch, Maria Dubovitskaya.