Privacy-Preserving Deep Learning

Shokri, Reza; Shmatikov, Vitaly

doi:10.1145/2810103.2813687

Cited by 1,559 publications

(1,188 citation statements)

References 38 publications

Supporting

Mentioning

1,142

Contrasting

Unclassified

Order By: Relevance

“…Beside these topics, there are many other examples: a methodology to estimate the uncertainty of the prediction that artificial intelligence model generates; principled way to integrate clinical or medical knowledge into the model training; development of content-based case retrieval system that searches images of similar diseases or conditions; efficient analysis of higherdimensional medical image, such as contrast enhanced images or follow-up images; and research on the privacy and security related to medical images when training and implementing artificial intelligence models [28,43,44,45,46]. These topics have not been thoroughly studied, and larger-scale studies are required.…”

Section: Discussionmentioning

confidence: 99%

Deep Learning for Medical Image Analysis: Applications to Computed Tomography and Magnetic Resonance Imaging

Jung¹,

Park²,

Hwang

2017

Hanyang Med Rev

View full text Add to dashboard Cite

Section: Discussionmentioning

confidence: 99%

Deep Learning for Medical Image Analysis: Applications to Computed Tomography and Magnetic Resonance Imaging

Jung¹,

Park²,

Hwang

2017

Hanyang Med Rev

View full text Add to dashboard Cite

“…The PS remains at a central location. In this context, we assume that the training data reside with the workers; this serves for instance as a basis for privacy-preserving scenarios [11], where users have their photos at home, and want to contribute to the computing of a global photo classification model, but without sending their personal data to a cloud service.…”

Section: Distributed Deep Learning On Edge-devicesmentioning

confidence: 99%

“…Meanwhile, the data to be processed at edge-devices themselves is not the limiting factor (3.9MB each in this experiment, as the dataset is split among workers). Our solution is to introduce a novel compression technique for sending updates from workers to the PS, using gradient selection [11]. We thus study the model accuracy with regards to update compression, as well as with regards to device reliability.…”

Section: Introductionmentioning

confidence: 99%

Distributed deep learning on edge-devices: Feasibility via adaptive compression

Hardy

Merrer

Séricola³

2017

2017 IEEE 16th International Symposium on Network Computing and Applications (NCA)

View full text Add to dashboard Cite

Abstract-A large portion of data mining and analytic services use modern machine learning techniques, such as deep learning. The state-of-the-art results by deep learning come at the price of an intensive use of computing resources. The leading frameworks (e.g., TensorFlow) are executed on GPUs or on high-end servers in datacenters. On the other end, there is a proliferation of personal devices with possibly free CPU cycles; this can enable services to run in users' homes, embedding machine learning operations. In this paper, we ask the following question: Is distributed deep learning computation on WAN connected devices feasible, in spite of the traffic caused by learning tasks? We show that such a setup rises some important challenges, most notably the ingress traffic that the servers hosting the up-to-date model have to sustain.In order to reduce this stress, we propose AdaComp, a novel algorithm for compressing worker updates to the model on the server. Applicable to stochastic gradient descent based approaches, it combines efficient gradient selection and learning rate modulation. We then experiment and measure the impact of compression, device heterogeneity and reliability on the accuracy of learned models, with an emulator platform that embeds TensorFlow into Linux containers. We report a reduction of the total amount of data sent by workers to the server by two order of magnitude (e.g., 191-fold reduction for a convolutional network on the MNIST dataset), when compared to a standard asynchronous stochastic gradient descent, while preserving model accuracy.

show abstract

“…Meanwhile, a data analyst who holds the global model will provide a computation result of ranking of any user's private data. Reference [2] also lets participants jointly learn a model while no input data sets of participants are revealed, but in training process participants should share small subaggregates of their models' key parameters. In the case of massive data stored in an untrusted server by a client, when the client would like to calculate a function on some part of its outsourced data, it could require the server.…”

Section: Related Workmentioning

confidence: 99%

Privacy Preserved Self-Awareness on the Community via Crowd Sensing

Fan

Xing

Tan

et al. 2017

Security and Communication Networks

View full text Add to dashboard Cite

In this method, firstly a polynomial approximation function model is set up for each user. Afterwards, "fragment" the coefficients of each model into pieces. Eventually "blend" all scraps to build the global model of all users. Users can use the global model to gain their corresponding ranking results after a special computing. Security analyses of three aspects elaborate the validity of proposed privacy method, even if some spiteful attackers try to steal private data of users, no matter who they are (users or someone outside the community). Experiments results manifest that the global model competently fits all users data and all privacy data are protected.

show abstract

Privacy-Preserving Deep Learning

Cited by 1,559 publications

References 38 publications

Deep Learning for Medical Image Analysis: Applications to Computed Tomography and Magnetic Resonance Imaging

Deep Learning for Medical Image Analysis: Applications to Computed Tomography and Magnetic Resonance Imaging

Distributed deep learning on edge-devices: Feasibility via adaptive compression

Privacy Preserved Self-Awareness on the Community via Crowd Sensing

Contact Info

Product

Resources

About