A Hilbert-Curve Based Encoding Scheme for Privacy-Preserving Nearest-Neighbor Classification

A Hilbert-Curve Based Encoding Scheme for Privacy-Preserving Nearest-Neighbor Classification

연구 분야: Analysis

학회: International Conference on Neural Information Processing

Usable privacy is a critical need of the present times. The encoded data in order to render usable privacy – i] should possess a certain degree of obfuscation for trespassers, ii] should be computable, iii] the computational result rendered from it (by the rightful entities) should not be much different from the one obtained from the original data. In this work, we address k-nearest neighbor (kNN) classification in a privacy-preserving framework where obtaining admissible computational efficacy (for the complete, original data) from the encoded database is the main focus. To this end, a Hilbert-curve-based feature encoding scheme is proposed. In our work, all computations are carried out in the encoded feature space (without invoking original data in any stages of the computation). The query point is also encoded before its interaction with the encoded database. Additionally, we propose a distance function to compute the separation of the encoded points. The empirical study on CIFAR10 and nine other real-world datasets shows that the kNN classification results on the encoded data obtained through the proposed method are comparable to that of the non-encoded, complete data. The performance of the proposed method is also superior to that of the two other privacy-preserving data obfuscation schemes included in the study. We also obtain a computationally admissible output from a fast, approximate kNN procedure on the obfuscated data.