過去的傳統神經網路搜索(Neural Architecture Search，NAS)方法減輕了一部分人工建構神經網路的負擔，但由於一般的自動化搜尋神經網路架構仍採用對不同架構訓練後觀察結果並調整架構再訓練，雖然自動化了手動設計架構的過程，進行實驗時還須龐大計算成本，並且無法及時應對邊緣設備的硬體限制。
本論文強化一個主從式架構的模型搜尋方法，透過預定義的超級網路和代理模型的應用，節省實際訓練的時間達到有效的網路推薦並保護使用者的資料隱私。透過事前給定資料運算限制或是硬體資訊來客製化最終推薦模型，以適應硬體的效能並維持一定的表現，進而解決傳統神經網路搜索無法應付的需求。除了提供計算中心的計算能力之外，本論文也透過主從式架構保護了使用者的資料隱私，並在保護資料隱私的同時透過回傳表現調整代理模型與搜索策略以應對不同的私有資料集樣貌。
最後本論文也做了兩階段的實驗來證明其有效性。第一個階段使用公開資料集訓練出收斂的代理模型，第二階段則透過主從式的資料傳輸使用私有資料集的表現校正代理模型的預測方向以及搜索策略的搜索方向，以達到遷移學習的結果。最後實驗的成績顯示我們可與state-of-the-art的NAS成果相匹配，並且成功在多種延遲時間限制下完成任務並達到一定的準確度。

In the past, Neural Architecture Search (NAS) mitigated a part of burden that deep learning scientists might suffer from. However, most of those methods usually take massive calculation resources to reach the results. To people who do the deep neural network task on edge or other IoTs barely be able to handle such huge calculations.
In this article, we proposed a distributed NAS method which handles the problems mentioned. Moreover, by take advantage of predefined supernetwork and surrogate models, we successfully save the NAS task time consuming and protect the security of privacy dataset at the same time. Last but not least, we borrow the calculation power of calculation center to deal with the most time-consuming part so that the users are able to finish their task at edge.
The proposed method can be divided into two parts: To retrieve convergence surrogate model and one best recommended model for public dataset, we have the first server-side pretraining. Second, client-side training which is meant to fetch the final best model for private dataset by using surrogate model and best recommended model for public dataset on training NAS. The proposed method is called Adaptive Portable Fast Platform-Aware Neural Architecture Search.

論文審定書 i
誌謝 ii
中文摘要 iii
Abstract iv
目錄 v
圖次 vii
表次 viii
第一章 緒論 1
1.1 研究背景與動機 1
1.2 問題描述 2
1.3 論文架構 3
第二章 文獻探討 4
2.1 自動化深度學習 5
2.1.1 NAS 5
2.1.2 HW-NAS 7
2.1.3 Transfer Learning in AutoDL 10
2.2 開源自動化學習工具 10
2.2.1 Auto-sklearn 11
2.2.2 Auto-keras 12
2.2.3 NNI 12
2.3 使用者訓練資料安全 13
2.3.1 Membership attack 14
2.3.2 Training data extraction 14
第三章 研究方法 15
3.1 資料集 15
3.1.1 ImageNet 15
3.1.2 CIFAR10/CIFAR100 15
3.1.3 MedMNISTv2 16
3.2 神經網路架構搜尋流程 19
3.2.1 資料傳遞步驟 19
3.2.2 元件介紹 24
第四章 實驗結果與分析 26
4.1 研究環境設置 26
4.1.1 硬體設備 26
4.1.2 軟體環境 26
4.2 實驗流程 27
4.2.1 伺服器端 29
4.2.2 客戶端 32
4.2.3 與PFP-NAS之比較 36
4.2.4 自動化工具與本論文比較 37
第五章 結論與未來展望 38
參考文獻 39
附錄一：口試委員提問 42
附錄二：無違反學術倫理聲明書 49
附錄三：相似度比對報告 51

[1] F. Rosenblatt, "Principles of neurodynamics. perceptrons and the theory of brain mechanisms," Cornell Aeronautical Lab Inc Buffalo NY, 1961.
[2] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, "Imagenet: A large-scale hierarchical image database," in 2009 IEEE conference on computer vision and pattern recognition, 2009: Ieee, pp. 248-255.
[3] O. Russakovsky et al., "Imagenet large scale visual recognition challenge," International journal of computer vision, vol. 115, no. 3, pp. 211-252, 2015.
[4] A. Krizhevsky, I. Sutskever, and G. E. Hinton, "Imagenet classification with deep convolutional neural networks," Communications of the ACM, vol. 60, no. 6, pp. 84-90, 2017.
[5] M. Tan et al., "Mnasnet: Platform-aware neural architecture search for mobile," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 2820-2828.
[6] B. Wu et al., "Fbnet: Hardware-aware efficient convnet design via differentiable neural architecture search," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 10734-10742.
[7] H. Cai, C. Gan, T. Wang, Z. Zhang, and S. Han, "Once-for-all: Train one network and specialize it for efficient deployment," arXiv preprint arXiv:1908.09791, 2019.
[8] K.-T. Ding, H.-S. Chen, Y.-L. Pan, H.-H. Chen, Y.-C. Lin, and S.-H. Hung, "Portable Fast Platform-Aware Neural Architecture Search for Edge/Mobile Computing AI Applications."
[9] Microsoft, "Neural Network Intelligence," 1 2021. [Online]. Available: https://github.com/microsoft/nni.
[10] X. Dong and Y. Yang, "Nas-bench-201: Extending the scope of reproducible neural architecture search," arXiv preprint arXiv:2001.00326, 2020.
[11] L. Xie and A. Yuille, "Genetic cnn," in Proceedings of the IEEE international conference on computer vision, 2017, pp. 1379-1388.
[12] P. Ren et al., "A comprehensive survey of neural architecture search: Challenges and solutions," ACM Computing Surveys (CSUR), vol. 54, no. 4, pp. 1-34, 2021.
[13] E. Real et al., "Large-scale evolution of image classifiers," in International Conference on Machine Learning, 2017: PMLR, pp. 2902-2911.
[14] M. Lin et al., "Zen-nas: A zero-shot nas for high-performance image recognition," in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 347-356.
[15] E. Jang, S. Gu, and B. Poole, "Categorical reparameterization with gumbel-softmax," arXiv preprint arXiv:1611.01144, 2016.
[16] B. Wu et al., "Fbnetv5: Neural architecture search for multiple tasks in one run," arXiv preprint arXiv:2111.10007, 2021.
[17] H. Cai, L. Zhu, and S. Han, "Proxylessnas: Direct neural architecture search on target task and hardware," arXiv preprint arXiv:1812.00332, 2018.
[18] A. Howard et al., "Searching for mobilenetv3," in Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 1314-1324.
[19] K. He, X. Zhang, S. Ren, and J. Sun, "Deep residual learning for image recognition," in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770-778.
[20] E. Real, A. Aggarwal, Y. Huang, and Q. V. Le, "Regularized evolution for image classifier architecture search," in Proceedings of the aaai conference on artificial intelligence, 2019, vol. 33, no. 01, pp. 4780-4789.
[21] C. Wong, N. Houlsby, Y. Lu, and A. Gesmundo, "Transfer learning with neural automl," Advances in neural information processing systems, vol. 31, 2018.
[22] M. Feurer, A. Klein, K. Eggensperger, J. Springenberg, M. Blum, and F. Hutter, "Efficient and robust automated machine learning," Advances in neural information processing systems, vol. 28, 2015.
[23] F. Pedregosa et al., "Scikit-learn: Machine learning in Python," the Journal of machine Learning research, vol. 12, pp. 2825-2830, 2011.
[24] M. Feurer, K. Eggensperger, S. Falkner, M. Lindauer, and F. Hutter, "Auto-sklearn 2.0: The next generation," arXiv preprint arXiv:2007.04074, vol. 24, 2020.
[25] H. Jin, Q. Song, and X. Hu, "Auto-keras: An efficient neural architecture search system," in Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, 2019, pp. 1946-1956.
[26] Q. Zhang et al., "Retiarii: A Deep Learning {Exploratory-Training} Framework," in 14th USENIX Symposium on Operating Systems Design and Implementation (OSDI 20), 2020, pp. 919-936.
[27] M. Rigaki and S. Garcia, "A survey of privacy attacks in machine learning," arXiv preprint arXiv:2007.07646, 2020.
[28] R. Shokri, M. Stronati, C. Song, and V. Shmatikov, "Membership inference attacks against machine learning models," in 2017 IEEE symposium on security and privacy (SP), 2017: IEEE, pp. 3-18.
[29] M. Fredrikson, S. Jha, and T. Ristenpart, "Model inversion attacks that exploit confidence information and basic countermeasures," in Proceedings of the 22nd ACM SIGSAC conference on computer and communications security, 2015, pp.
1322-1333.
[30] A. Krizhevsky and G. Hinton, "Learning multiple layers of features from tiny images," 2009.
[31] J. Yang, R. Shi, and B. Ni, "Medmnist classification decathlon: A lightweight automl benchmark for medical image analysis," in 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI), 2021: IEEE, pp. 191-195.
[32] S. Kirkpatrick, C. D. Gelatt Jr, and M. P. Vecchi, "Optimization by simulated annealing," science, vol. 220, no. 4598, pp. 671-680, 1983.