Privacy-Preserving Neural Representation for Brain-Inspired Learning

Javier Roberto Rubalcava-Cortés, Alejandro Hernández-Cano, Alejandra Citlalli Pacheco-Tovar, Farhad Imani, Rosario Cammarota, Mohsen Imani

Abstract:

In this paper, we propose BIPOD, a brain-inspired
privacy-oriented machine learning. Our method rethinks privacy-
preserving mechanisms by looking at how the human brain
provides effective privacy with minimal cost. BIPOD exploits
hyperdimensional computing (HDC) as a neurally-inspired com-
putational model. HDC is motivated by the observation that the
human brain operates on high-dimensional data representations.
In HDC, objects are thereby encoded with high-dimensional
vectors, called hypervectors, which have thousands of elements.
BIPOD exploits this encoding as a holographic projection with
both cryptographic and randomization-based features. BIPOD
encoding is performed using a set of brain keys that are generated
randomly. Therefore, attackers cannot get encoded data without
accessing the encoding keys. In addition, revealing the encoding
keys does not directly translate to information loss. We enhance
BIPOD encoding method to mathematically create perturbation
on encoded neural patterns to ensure a limited amount of infor-
mation can be extracted from the encoded data. Since BIPOD
encoding is a part of the learning process, thus can be optimized
together to provide the best trade-off between accuracy, privacy,
and efficiency. Our evaluation on a wide range of applications
shows that BIPOD privacy-preserving techniques result in 11.3×
higher information privacy with no loss in classification accuracy.
In addition, at the same quality of learning, BIPOD provides
significantly higher information privacy compared to state-of-
the-art privacy-preserving techniques.