Differentially Private Compressive K-means

Schellekens, Vincent; Chatalic, Antoine; Houssiau, Florimond; Montjoye, Y.-A. de; Jacques, Laurent; Gribonval, Rémi

doi:10.1109/icassp.2019.8682829

Cited by 14 publications

(8 citation statements)

References 24 publications

(28 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…An interesting perspective is to characterize such schemes in terms of tradeoffs between differential privacy and ability to learn from the resulting sketch. Preliminary results in this direction have been recently achieved Schellekens et al [2019].…”

Section: Discussionmentioning

confidence: 98%

Compressive statistical learning with random feature moments

et al. 2021

View full text Add to dashboard Cite

Section: Discussionmentioning

confidence: 98%

Compressive statistical learning with random feature moments

et al. 2021

View full text Add to dashboard Cite

“…The NSR was indeed shown empirically [45] to be a good proxy to estimate the utility of a sketching mechanism for the task of clustering where performance is measured with the SSE (sum of squared errors) defined in (2). Figures 4 and 5 give an overview of this correlation.…”

Section: Noise To Signal Ratio As a Proxy For Utilitymentioning

confidence: 99%

“…A first and reduced version of this work with privacy upper bounds and without the subsampling mechanism has been previously published[45].…”

mentioning

confidence: 99%

Compressive learning with privacy guarantees

Chatalic¹,

Schellekens²,

Houssiau

et al. 2021

Information and Inference: A Journal of the IMA

View full text Add to dashboard Cite

This work addresses the problem of learning from large collections of data with privacy guarantees. The compressive learning framework proposes to deal with the large scale of datasets by compressing them into a single vector of generalized random moments, called a sketch vector, from which the learning task is then performed. We provide sharp bounds on the so-called sensitivity of this sketching mechanism. This allows us to leverage standard techniques to ensure differential privacy—a well-established formalism for defining and quantifying the privacy of a random mechanism—by adding Laplace of Gaussian noise to the sketch. We combine these standard mechanisms with a new feature subsampling mechanism, which reduces the computational cost without damaging privacy. The overall framework is applied to the tasks of Gaussian modeling, k-means clustering and principal component analysis, for which sharp privacy bounds are derived. Empirically, the quality (for subsequent learning) of the compressed representation produced by our mechanism is strongly related with the induced noise level, for which we give analytical expressions.

show abstract

“…Another common strategy for private clustering is based on differential privacy (DP) [6,13,18,22]. DP consists on adding "statistical noise" that is significant enough to protect client's privacy, but small enough to not affect the model performance.…”

Section: Privacy Preserving Distributed K-meansmentioning

confidence: 99%

Efficient Privacy Preserving Distributed K-Means for Non-IID Data

Brandão

Mendes

Vilela

2021

Advances in Intelligent Data Analysis XIX

View full text Add to dashboard Cite

Privacy is becoming a crucial requirement in many machine learning systems. In this paper we introduce an efficient and secure distributed K-Means algorithm, that is robust to non-IID data. The base idea of our proposal consists in each client computing the K-Means algorithm locally, with a variable number of clusters. The server will use the resultant centroids to apply the K-Means algorithm again, discovering the global centroids. To maintain the client's privacy, homomorphic encryption and secure aggregation is used in the process of learning the global centroids. This algorithm is efficient and reduces transmission costs, since only the local centroids are used to find the global centroids. In our experimental evaluation, we demonstrate that our strategy achieves a similar performance to the centralized version even in cases where the data follows an extreme non-IID form.

show abstract

Differentially Private Compressive K-means

Cited by 14 publications

References 24 publications

Compressive statistical learning with random feature moments

Compressive statistical learning with random feature moments

Compressive learning with privacy guarantees

Efficient Privacy Preserving Distributed K-Means for Non-IID Data

Contact Info

Product

Resources

About