Towards Privacy-Preserving Intelligence: Differential Privacy in Machine Learning

Overview

Western Connecticut State University Although scary, data collection is inevitable. Our society needs data to improve. But, unlike popular opinion, data collection does not necessarily entail privacy loss. For decades, data-holding organizations have anonymized public datasets, removing identifying information such as names. However, hackers can now link two public datasets containing the same person and find who that person is—thus compromising privacy. Differential privacy prevents these linkage attacks by using computer mechanisms to obscure data with randomized numbers (noise). Mechanisms add just enough noise to datasets to make individuals undetectable. However, current mechanisms are not complex enough: they neglect next-generation hacking techniques and data reconstruction through noise averaging. Overly-secure mechanisms, on the other hand, add too much noise, preventing analysts from learning from the data. To fix these issues, I create three differentially private mechanisms that first use the discrete M- band wavelet transform, which preserves the energy of the data. The mechanisms LS and LS+ then use a “Laplace- Sigmoid” distribution that multiplies Laplace-distributed values with the sigmoid function, creating a doubly random distribution to draw noise from. The third mechanism utilizes pseudo-quantum steganography, which simulates qubit encryption, to embed noise into data. I then test the mechanisms in five machine learning environments. The mechanisms achieve more than 94% classification accuracy for all privacy values tested, proving that they successfully retain both differential privacy and statistical learnability. As data privacy becomes exigent and quantum computing emerges, my research links the two branches and portrays what data privacy could look like in the future.

Competition history

• JSHS 2020 Category not listed

Resources

• National JSHS abstract book ↗

Related projects

Closest projects by meaning, across every fair and year in the corpus.