D5.3 - Implementation of Provisioning, Outsourcing and Processing Frameworks

This report overviews the complete specification and the evaluation of four different TREDISEC primitives that enable a cloud server to process data while not being able to access its content. Thanks to these primitives, cloud customers can delegate their computationally expensive operations to the untrusted cloud server and hence benefit from the performance advantages offered by this new technology while preserving data privacy. These four primitives are the following:

Configuration tool for privacy preserving Data provisioning and Outsourcing

We propose a configuration management tool to help security administrators securely outsource their encrypted relational SQL database into the untrusted cloud server. During the migration of the data, this tool analyses all utility and security constraints, detects conflicts between them, if any, and offers some conflict resolution procedures. This new tool improves the performance and effectiveness of the entire data provisioning and outsourcing framework described in previous deliverables D5.1 and D5.2.

Multi-User Searchable Encryption

This primitive addresses the problem of searchable encryption in the multi-user context whereby multiple users outsource their encrypted files and the corresponding indices into the cloud and allow multiple other users to query them while not revealing any sensitive information to the cloud server. This primitive was previously introduced in D5.2. In this deliverable we provide its complete specification and evaluate its security and performance. We also identify a serious privacy leakage that almost all existing multi-user searchable encryption solutions suffer from and define a dedicated security model that needs to be taken into account during the design of a new multi-user searchable encryption scheme.

Authenticated Encryption

Authenticated encryption (AE) is a symmetric-key encryption mechanism that in addition to confidentiality also delivers integrity and authenticity. It has been shown in the literature that AE solutions can be considered as essential building blocks for verifiable searchable encryption solutions as the one that were presented in the previous deliverable (D5.2). In this deliverable, we propose to improve the effectiveness of existing AE schemes by supporting variable-length tags per key for ciphertext expansion without sacrificing functional and security properties. To this end, we provide a formal definition for the notion of nonce-based variable-stretch AE (nvAE) and further propose some extensions to existing solutions to securely achieve this property.

Privacy preserving feature extraction for Biometrics

This new primitive enables a cloud customer to outsource a classification algorithm in the context of deep neural networks. We assume that the cloud customer already obtained a trained neural network model and would like to outsource the classification operation to the cloud while keeping the underlying sensitive information private. Indeed, the cloud should neither discover the input to the classification algorithm not its output (the resulting classification label). To achieve the privacy requirement, the proposed solution uses a fully homomorphic encryption (FHE) scheme while keeping the number of multiplication operations low. The underlying algorithmic building block (namely the ReLU activation function) is therefore approximated into a low degree polynomial which makes the neural network compatible with FHE and at the same time without much losing its accuracy.