Logo image
Back
Secure signal processing and secure machine learning using fully homomorphic encryption
Dissertation   Open access

Secure signal processing and secure machine learning using fully homomorphic encryption

Thomas M. Shortell
Doctor of Philosophy (Ph.D.), Drexel University
Oct 2018
DOI:
https://doi.org/10.17918/D8D10G
pdf
Shortell_Thomas_20182.82 MBDownloadView
PDF Open Access Open Access (License Unspecified)

Abstract

Data encryption (Computer science) Computer Science Machine Learning Signal Processing
This dissertation focuses on the new techniques for secure and private computation for signal processing and machine learning. Specifically, the thesis will focuses on extending Fully Homomorphic Encryption (FHE) technique in a cloud computing set up by running the algorithms while the data is encrypted. The (FHE) comes at a cost of integer space and not the real space needed by signal processing and machine learning algorithms. Solving this problem requires using numerical models that represent real numbers in an integers space including a rational number format and a fixed point binary format. These models allow the computation of signal processing and machine learning algorithms while the data is encrypted. This dissertation includes analysis and implementation of a natural logarithm, Brightness-Contrast filter, Fast Fourier Transform, Speeded Up Robust Features, Histogram of Oriented Gradients, and Convolutional Neural Networks. Analyzing these algorithms with the numerical models provide tight upper-bounds on numerical error introduced from their use. Each of the implementations provide unique understanding of error propagation in the encrypted domain. Experimental results of each implementation were aligned with the expected error based on the theorems. Despite their algorithmic constructions, each implementation is a step towards more advanced computations for privacy and security in the cloud.

Metrics

73 File views/ downloads
48 Record Views

Details

Logo image