Inferential Separation for Privacy: Irrelevant Statistics and Quantization

This work presents a new paradigm for protection of sensitive inferences drawn from data streams with relevance to Internet-of-Things (IoT). This paradigm is an alternative to end-to-end encryption of entire data streams, or noise-addition based privatization mechanisms. It relies on the notion that raw data shared through IoTs are themselves not sensitive but for the inferences that can be drawn from them, and further, these inferences vary much slower than the collected data. Methodologies are developed that transform data streams into two parallel sub-streams of minimum sufficient and maximal irrelevant statistics, such that the sparse minimal sufficient stream can be protected using encryption, and the high rate irrelevant stream is guaranteed to provide perfect privacy for the underlying inference without any additional protection. This inferential separation is explored theoretically, where it is proved that the inference relevant (minimum sufficient) stream grows as
$O(\log t)$
for a data stream of length
$t$
. The approach is extended to bandwidth constrained devices, where a new optimal quantization scheme is presented that achieves maximum fidelity while guaranteeing privacy. The presented algorithms are demonstrated to practical IoT datasets where trained CNN based classifiers are shown to fail on the unprotected high rate stream.