臺灣博碩士論文加值系統

生成對抗網路除了生成仿真資料外，也常拿來進行資料域集的轉換，本論文應用這個機制的精神，設計一個半監督式神經網路訓練方法，將資料中所有者的識別資訊遮蔽起來，維持資料集原本的主要用途，例如本論文中維持手機蒐集的三軸加速度資料的動作分辨率下，盡量降低可識別資料提供者的相關資訊，以達到多媒體資料去識別化的目的，經特徵工程處理過後的561維度三軸加速度資料，身分辨識率可由75%大幅下降至20%，動作辨識仍可維持98.5%的高辨識率；若是對時序三軸加速度資料進行去識別化，身分辨識率可由99%大幅下降至20%，動作辨識仍可維持94%的高辨識率，且對於未參與訓練的測試資料也具備一般化的能力，未參與去識別化訓練的561維度三軸加速度資料身分辨識率從81.8%降到35%，未參與去識別化訓練的時序三軸加速度資料身分辨識率從99.8%降到45.6%。

Generative Adversarial Networks are often used for very photorealistic data generation, style transfer, and feature distentanglement. In this paper, a semi-supervised deep-learning algorithm is designed for the de-identification of large datasets. The owner can process the dataset for hiding the identity-related information, while the dataset could still be useful for its original purposes.
In this paper, the dataset consisting of three-axis accelerometer and three-axis gyroscope data captured from smartphone for human activity recognition is used as an example for the de-identification experiments. There are two forms of the data: the first consists of pre-processed 561-feature vectors and the second consists of raw time sequences. For the first one, the high accuracy of activity recognition is maintained as 98.5%, while the accuracy of human identification rate drops from 75% to 20%. For the second one, the accuracy of activity recognition is 94% and the accuracy of human identification drops from 99.25% to 20%. Even for those data which are not used in training the de-identification network, the corresponding human activities of the de-identified version still can be recognized correctly. It shows that the identification network has good generalization capability.

目次

摘要 I
Abstract II
目次 III
圖次 V
表次 VI
第一章 前言 1
1.1文獻回顧 2
1.2 本文貢獻 3
1.3章節內容介紹 3
第二章 背景知識 5
2.1 深度學習(Deep learning) 5
2.1.1損失函數(Loss function) 5
2.1.2誤差反向傳播(Backpropagation, BP) 5
2.2 一維卷積神經網路(1D convolution Neural Network, 1DCNN) 6
2.3 生成對抗網路(Generative adversarial Network, GAN) 8
2.3 三軸資料 9
2.3.1 時序三軸資料 9
2.3.2 處理過之三軸加速度資料 10
第三章 去識別化系統 11
3.1 目標資料集 11
3.2 去識別化 11
3.3 去識別化生成對抗網路 11
3.3.1 去識別化網路 12
3.3.2 身分判別器網路 12
3.3.3 動作分類網路 12
第四章 訓練方法 14
4.1 資料和標籤 14
4.2 模型間的關係 14
4.3 損失函數 15
4.5 訓練演算法 15
第五章 系統檢測 17
5.1 動作分類檢測 17
5.2 身分識別檢測 17
5.3 一般化檢測 18
5.3.1 一般化實驗資料分配 18
5.3.2 一般化動作測試 19
5.3.3 一般化身分測試 19
第六章 時序三軸資料去識別化 21
6.1 三軸時序資料重疊取樣(overlap) 21
6.2 時序資料的分辨 21
6.3 去識別化網路的修改 22
6.4 去識別化後動作分辨 22
6.4 去識別化後身分辨識 22
6.5時序資料的混淆矩陣 23
6.6其他方法測試 24
第七章 結論 26
參考文獻 27

[1] E. Alpaydin, “Introduction to Machine Learning,” The MIT Press2014.
[2] D. Anguita, A. Ghio, L. Oneto, X. Parra, and J.L. Reyes-Ortiz, “Human Activity Recognition on Smartphones Using a Multiclass Hardware-Friendly Support Vector Machine,” Springer Berlin Heidelberg, Berlin, Heidelberg, 2012, pp. 216-223.
[3] M.W. Gardner, and S. Dorling, “Artificial neural networks (the multilayer perceptron)—a review of applications in the atmospheric sciences,” Atmospheric environment, 32,1998, pp.2627-2636.
[4] X. Glorot, and Y. Bengio, “Understanding the difficulty of training deep feedforward neural networks,” Proceedings of the thirteenth international conference on artificial intelligence and statistics, 2010, pp. 249-256.
[5] X. Glorot, A. Bordes, and Y. Bengio, “Deep sparse rectifier neural networks,” Proceedings of the fourteenth international conference on artificial intelligence and statistics, 2011, pp. 315-323.
[6] I.J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, “Generative adversarial nets,” Proceedings of the 27th International Conference on Neural Information Processing Systems - Volume 2, MIT Press, Montreal, Canada, 2014, pp. 2672-2680.
[7] C.J. Hoofnagle, B. van der Sloot, and F.Z. Borgesius, “The European Union general data protection regulation: what it is and what it means,” Information & Communications Technology Law, 28,2019, pp.65-98.
[8] A. Krizhevsky, I. Sutskever, and G.E. Hinton, “Imagenet classification with deep convolutional neural networks,” Advances in neural information processing systems, 2012, pp. 1097-1105.
[9] B. Meden, Ž. Emeršič, V. Štruc, and P. Peer, “k-Same-Net: k-Anonymity with Generative Deep Neural Networks for Face Deidentification,” Entropy, 20,2018, pp.60.
[10] D.E. Rumelhart, G.E. Hinton, and R.J. Williams, “Learning representations by back-propagating errors,” Nature, 323,1986, pp.533-536.
[11] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov, “Dropout: a simple way to prevent neural networks from overfitting,” The journal of machine learning research, 15,2014, pp.1929-1958.
[12] L. Sweeney, “k-anonymity: A model for protecting privacy,” International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, 10,2002, pp.557-570.
[13] C. Zhang, S. Bengio, M. Hardt, B. Recht, and O. Vinyals, “Understanding deep learning requires rethinking generalization,” International Conference on Learning Representations, Palais des Congrès Neptune, Toulon, France, 2017.