Oblivious Neural Network Computing via Homomorphic Encryption

By Claudio Orlandi, Alessandro Piva, Mauro Barni. 2007

143 citations recorded by the source corpus.

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

Citations in the corpus

Secure Evaluation of Quantized Neural Networks (2019, methodology): “like language and so it is easy to extend it with support for more models in the future. D. Related Work Secure evaluation of Neural Networks can be traced back to at least the work by Orlandi et al. [5, 51] which present a solution based on HE techniques. Several later works rely on HE techniques e”
Distributed learning of deep neural network over multiple agents (2018, methodology): “Some interesting solutions include using Ada-boost to jointly train classiﬁer ensembles [28], using random rotation perturbations for homomorphic pseudo-encryption [29] and applying homomorphic cryptosystem to perform secure computation [30].”
DeepSecure: Scalable Provably-Secure Deep Learning (2017, background): “quires the data owner to have a constant connection with the cloud server while evaluating the DL network. To address the issue of information leakage as a result of sharing the intermediate results, [39] and [40] enhance the protocol initially proposed in [38] to obscure the weights. However, even ”
Chameleon: A Hybrid Secure Computation Framework for Machine Learning Applications (2018, background): “ toexecuteeachlayerinChameleon,includingthenecessary protocol conversions. FurtherRelatedWorks.One of the earliest solutions for obliviously evaluating a neural network was proposed by Orlandi et al. [75]. They suggest adding fake neurons to the hidden layers in 10 the original network and evaluatin”
Cheetah: Optimizing and Accelerating Homomorphic Encryption for Private Inference (2020, background): “MPC-based schemes provide an alternative to approximation by combining HE with other security solutions, typically garbled circuit (GC) [33], [37], [42], [46], [50], [51].”
Privacy-Preserving Classification on Deep Neural Network (2017, background): “Orlandi et al [36] enhanced the protocol [2]: intermediate results are no longer revealed to the client.”
Privacy-Preserving ECG Classification With Branching Programs and Neural Networks (2011, background): “, [43], where the activation functions are implemented by introducing a rather security-critical multiplicative blinding step), our solution is provably secure and computationally efficient.”
Secure Outsourced Matrix Computation and Application to Neural Networks (2018, unknown): “The second scenario was considered by Makri et al. [35] based on Multi-Party Computation (MPC) using polynomial kernel support vector machine classification.”
No Peek: A Survey of private distributed deep learning (2018, methodology): “…include de-identiﬁcation methods like anonymization [52], obfuscation methods like diﬀerential privacy [100,101,102] and cryptographic techniques like homomorphic encryption [19,93,96] and secure multi-party computation (MPC) protocols like oblivious transfer [84,47] and garbled circuits [41].”
Oblivious Neural Network Predictions via MiniONN Transformations (2017, result): “[47] noticed that this process leaks significant information about S’s neural network, and proposed a method to obscure the intermediate results.”
XONN: XNOR-based Oblivious Deep Neural Network Inference (2019, background): “oblivious inference is proposed by Barni et al. [38] that leverages homomorphic encryption. Howerver, this work leaks some information about the intermediate states of the computation. Orlandi et al. [39] improve upon this work and avoid any information leakage. In [40], authors propose a mechanism ”
Security and Privacy Issues in Deep Learning (2018, background): “[100] Shokri and Shmatikov [29] Liu et al.”
Towards Deep Neural Network Training on Encrypted Data (2019, background): “An early work along these lines is due to Barni et al. and Orlandi et al. [2, 25], that combined additively-homomorphic encryption with an interactive protocol, and were able to run the inference part of a small network in about ten seconds.”
EPIC: Efficient Private Image Classification (or: Learning from the Masters) (2019, methodology): “[36], also considers NN-based data classification extending the work of Barni et al.”