On Extreme Learning Machine For Preserving Privacies

As the development of the information technique, we pay more and more attention to the protection of private information while using the information services. Especially, in the procedure of statistical analysis and classification for large scale dataset, it has significant practical importance and is a huge technical challenge to preserve the privacy of individual while releasing the statistical property of the database as a whole.Differential privacy is one of the most promising privacy preserving techniques in the statistical query processing field at present. It prevents the adversary from inferring the whether any individual record in the database or not by injecting random noise into each query. The Extreme Learning Machine (ELM) is a learning algorithm mainly used for classification and regression. It can achieve much better generalization performance and much fast learning speed than the Support Vector Machine, which is the mostly extensively used technique for classification at present. Therefore, we pay a special attention to the research of ELM technique which supports privacy preserving to take account of the publication of statistical information and privacy preserving of individual information.This thesis surveys the basic privacy preserving techniques and the basic classification technique used to construct supporting privacy preserving classifier, and then constructs the privacy preserving extreme learning machine. In order to prevent the disclosure of privacy from classifier, this thesis gives a strong privacy attack model firstly, and gives a metric of privacy preserve of classifier to compare the privacy preserving capability of different privacy preserving classifiers. Then this thesis gets the equivalent Empirical Risk Minimization (ERM) form of ELM, and gives the proof for differential privacy capability of Differential Privacy Extreme Learning Machine (DPELM) by the proof for differential privacy empirical risk minimization. What’s more, to maintain the faster learning speed, the looser constrains and the better generalization performance, we build two different DPELMs, the output perturbation based DPELM and the object perturbation based DPELM in different ways.At last, this thesis conducts general experiments on the real datasets. Firstly, we validate the equivalence between ELM trained by linear system and its equivalent ERM form. Then we validate that DPELM can achieve well privacy preserving capability and maintain the faster learning speed, the well generalization performance, and it can achieve a better generalization performance than Differential Privacy Support Vector Machine under the same privacy request.