Compressive Privacy: From Information\/Estimation Theory to Machine Learning [Lecture Notes]

Privacy-Preserving Machine Learning: Threats and Solutions

“…While some entities might seek total hiding of their data, DR has another benefit for privacy. For datasets that have samples with two labels: a utility label and a privacy label, Kung 26 proposes a DR method to enable the data owner to project her data in a way that enables maximizing the accuracy of learning for the utility labels, while decreasing the accuracy for learning the privacy labels. Although this method does not eliminate all privacy risks of the data, it enables controlling the misuse of the data when the privacy target is known.…”

A compressive multi-kernel method for privacy-preserving machine learning

“…However, as the application of privacy is under consideration, extending the embedded feature space can come at a cost of extra information shared. Under the regime of Compressive Privacy [16], [17], this is not desirable. Hence, in the design of the multi-kernel step, the rank of the final multi-kernel should not exceed that of the original data.…”

“…3) Signal-to-Noise Ratio for Kernel Weight Design: The design of the multi-kernel in Equation ( 15) involves the determination of the value of µ l for each kernel. The signalto-noise ratio (SNR) in the form of the trace-norm of the discriminant matrix, as motivated by the inter-class separability metric in Equation ( 31) of [16], is proposed as a new metric to decide the value of µ l . Specifically, the SNR is defined as,…”

“…Fortunately, the SNR score can be extended to such setting by considering the utility-to-privacy ratio similar to that in [35], [16]. By utilizing both the utility label y u and privacy label y p , the utility-to-privacy ratio can be defined as,…”

“…More importantly, given L u classes of utility classification, DCA only requires L u −1 dimensions to reach the maximum discriminant power, permitting high compression when L u is much less than the original dimension of the data. The multi-kernel step, meanwhile, employs the nonnegative linear combination of kernels, using a novel nonuniform weighting method based on the signal-to-noise ratio (SNR) score of each kernel, as inspired by the inter-class separability metric presented in [16].…”