Loading...

Deep Feature Extraction by Information Regularization

Osia, Ali | 2019

827 Viewed
  1. Type of Document: Ph.D. Dissertation
  2. Language: Farsi
  3. Document No: 53122 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Rabiee, Hamid Reza
  7. Abstract:
  8. Feature Extraction is an inseparable part of many machine learning algorithms. The extracted feature from the input data, while preserving the label information as much as possible, should make label inference more feasible than the input data itself. Additionally, in many machine learning, information theory, and privacy problems we have another constraint on eliminating sensitive, confounding, or unwanted information from the feature. That is why the problem of feature extraction with information regularization (preserving desired information and removing unwanted one) becomes an important issue that has been considered from different aspects, inter-connecting various fields.In this dissertation, using an information theoretic approach, we model the problem of feature extraction with information regularization as an optimization problem. In order to find a local optimum solution, we propose a deep learning based framework which we use in two different applications: (I) preserving the privacy of test data in using cloud-based machine learning services; and (II) batch effect correction in image-based cell profiling. Moreover, we present a novel measure of privacy based on guessing entropy, which we use to evaluate information leakage of different models. Conducted experiments and their results demonstrate the effectiveness of the proposed framework in different applications. Finally, we define guessing leakage based on guessing entropy and find the optimal utility-privacy trade-off against brute-force adversary
  9. Keywords:
  10. Machine Learning ; Information Theory ; Deep Neural Networks ; Privacy ; Brute Force Attack ; Batch Effect Correction ; Guessing Entropy

 Digital Object List

 Bookmark

...see more