學術產出-Theses
Article View/Open
Publication Export
-
題名 階層式深度神經網路及其應用
Deep Hierarchical Neural Networks and Its Applications作者 朱家宏
Chu, Jia-Hong貢獻者 廖文宏
Liao, Wen-Hung
朱家宏
Chu, Jia-Hong關鍵詞 深度學習
神經網路
階層式分類
對抗例攻擊
Deep learning
Neural network
Hierarchical classification
Adversarial attack日期 2023 上傳時間 6-Jul-2023 17:03:54 (UTC+8) 摘要 綜觀圖形分類的應用,有若干的資料集在本質上都有著階層的概念,例如動 物辨識屬於物種分類,在標記上就有科、屬、種的階層關係,或是在衛星圖像中 的船隻分類,在大類別民用與軍用設施之下,分別有更細緻的分類,如民用的子 類別有漁船、遊艇,而軍用的子類別則有巡防艦、驅逐艦等。對於這些應用,我 們應該要設計出一個階層式模型,將影像輸入到此模型後,該模型會預測出階層 的分類答案出來,並且由於其為階層式分類,預測出的答案除了皆需保持階層依 賴關係一致性(預測出的細類別要為預測出的粗類別其一子類),也要避免模型 預測出階層上下雖有依賴關係,但其兩者都預測錯誤,造成嚴重的誤判性。所以 在量測階層式準確率時,應使用特定的指標來評估此模型的表現,而不是個別計 算各階層的 Top-1 準確率。準此,本論文提出三個階層式指標(綜合正確率、階 層關係一致性、預測風險率)來評估兩者階層式模型的預測好壞及預測錯誤時的 風險嚴重程度。在網路架構部分,本論文提出一個新的階層式架構模型及其訓練方法 HCN(Hierarchy-constrained Network),具體而言,該階層式模型在進行訓練時, 會將粗分類的特徵層融合到細分類的特徵層,同時也將模型的目標函數添加兩種 限制,第一種是限制粗細分類的特徵層,假設兩組影像同屬同一組粗分類,則這 兩張圖的這些特徵層數值要接近; 第二種則是模型預測出的粗分類及細分類類 別輸出需要保有一致關係(e.g. 細分類的狗就要對應到粗分類的動物而不是卡 車)。透過此架構與目標函數的設計,並測試在三種資料集上 (CIFAR100\\TinyImageNet-200\\CUB200-2011),我們發現此架構除了在階層式指 標有優良的表現外,因其訓練方式會使得同屬粗類別的細類別特徵層彼此數值更 相近,模型在受到對抗例攻擊時會較難將輸入影像的細分類判斷成與此影像不同 粗分類的細分類,以此來抵抗對抗例攻擊。
Hierarchical relationships exist in many datasets that involve object classification tasks. For example, animal recognition involves species classification, which is organized hierarchically into family, genus, and species. Similarly, ship classification in satellite images has finer sub-classes under main categories such as civilian and military vessels, including fishing boats, yachts, patrol ships, and destroyers. Therefore, it is crucial to design models capable of predicting hierarchical classification results.In hierarchical classification tasks, predicted answers should not only maintain hierarchical consistency (i.e., the predicted sub-class should be a child class of the predicted parent class), but also avoid serious misjudgments caused by incorrect predictions of both parent and child classes. Hence, it is important to design appropriate metrics to evaluate the performance of hierarchical models beyond calculating Top-1 accuracy for individual classes in each hierarchy. To address this issue, three new metrics, namely Aggregated Accuracy, Hierarchy Consistency, and Risk Factor, are proposed to accurately assess the performance of hierarchical models and the severity of prediction errors.In terms of network architecture, this thesis introduces a novel hierarchical model and training method called HCN (Hierarchy-constrained Network). During training, the features of the parent class are fused into the features of the child class. Additionally, two constraints are added to the objective function of the model. The first constraint considers the similarity of features between parent and child classes. If two sets of images belong to the same parent class, their features should be similar. The second constraint ensures that the predicted outputs of parent and child class categories maintain consistent relationships, such as the dog category of the child class corresponding to the animal category of the parent class, and not a truck category.Experimental analysis on three datasets, namely CIFAR100, TinyImageNet-200, and CUB200-2011, demonstrates that the proposed HCN outperforms existing hierarchical models based on the proposed hierarchical metrics. Furthermore, the features of child classes belonging to the same parent class are found to be more similar, making it difficult for the model to misclassify the sub-class of an input image into a sub-class of a different parent when subjected to adversarial attacks, thereby enhancing the overall robustness of the model.參考文獻 [ 1 ] Roy, D., Panda, P., & Roy, K. Tree-CNN: a hierarchical deep convolutional neural network for incremental learning. Neural Networks, 121, 148-160. 2020.[ 2 ] Jiang, S., Xu, T., Guo, J., & Zhang, J. Tree-CNN: from generalization to specialization. EURASIP Journal on Wireless Communications and Networking, 2018(1), 216. 2018.[ 3 ] Xinqi Zhu, Michael Bain. B-CNN: Branch Convolutional Neural Network for Hierarchical Classification. arXiv 2017, arXiv:1709.09890.2017.[ 4 ] Salma Taoufiq , Balázs Nagy , Csaba Benedek.HierarchyNet: Hierarchical CNN- Based Urban Building Classification. Remote Sensing,Volume 12 ,Issue 22 .2020.[ 5] Yan, Z., Zhang, H., Piramuthu, R., Jagadeesh, V., DeCoste, D., Di, W., & Yu, Y.. HD-CNN: hierarchical deep convolutional neural networks for large scale visual recognition. In Proceedings of the IEEE international conference on computer vision (pp. 2740-2748). 2015.[ 6 ] M. D. Zeiler and R. Fergus. Visualizing and Understanding Convolutional Networks, pages 818–833. Springer Interna- tional Publishing, Cham, 2014.[ 7 ] A. Krizhevsky and G. Hinton. Learning multiple layers of features from tiny images. Computer Science Department, University of Toronto, Tech. Rep, 2009.[ 8 ] Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., & Fei-Fei, L. Imagenet: A large- scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition (pp. 248–255).2009.[ 9 ] Tan, M. and Le, Q.V. EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. Proceedings of the 36th International Conference on Machine Learning, ICML 2019, Long Beach, 9-15 June 2019, 6105-6114, 2019.[ 10 ] D. Arthur and S. Vassilvitskii. k-means++: The advantages of careful seeding. In SODA, pages 1027–1035, 2007.[ 11 ] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The caltech-ucsd birds-200-2011 dataset. 2011.[ 12 ] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In CVPR, 2016.[ 13 ] Deng J, Guo J, Xue N, Zafeiriou SJAPA .Arcface: Additive angular margin loss for deep face recognition, 2018.[ 14 ] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In International Conference on Learning Representations (ICLR), 2015.[ 15 ] Francesco Croce, Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks, 2020. 描述 碩士
國立政治大學
資訊科學系
110753112資料來源 http://thesis.lib.nccu.edu.tw/record/#G0110753112 資料類型 thesis dc.contributor.advisor 廖文宏 zh_TW dc.contributor.advisor Liao, Wen-Hung en_US dc.contributor.author (Authors) 朱家宏 zh_TW dc.contributor.author (Authors) Chu, Jia-Hong en_US dc.creator (作者) 朱家宏 zh_TW dc.creator (作者) Chu, Jia-Hong en_US dc.date (日期) 2023 en_US dc.date.accessioned 6-Jul-2023 17:03:54 (UTC+8) - dc.date.available 6-Jul-2023 17:03:54 (UTC+8) - dc.date.issued (上傳時間) 6-Jul-2023 17:03:54 (UTC+8) - dc.identifier (Other Identifiers) G0110753112 en_US dc.identifier.uri (URI) http://nccur.lib.nccu.edu.tw/handle/140.119/145938 - dc.description (描述) 碩士 zh_TW dc.description (描述) 國立政治大學 zh_TW dc.description (描述) 資訊科學系 zh_TW dc.description (描述) 110753112 zh_TW dc.description.abstract (摘要) 綜觀圖形分類的應用,有若干的資料集在本質上都有著階層的概念,例如動 物辨識屬於物種分類,在標記上就有科、屬、種的階層關係,或是在衛星圖像中 的船隻分類,在大類別民用與軍用設施之下,分別有更細緻的分類,如民用的子 類別有漁船、遊艇,而軍用的子類別則有巡防艦、驅逐艦等。對於這些應用,我 們應該要設計出一個階層式模型,將影像輸入到此模型後,該模型會預測出階層 的分類答案出來,並且由於其為階層式分類,預測出的答案除了皆需保持階層依 賴關係一致性(預測出的細類別要為預測出的粗類別其一子類),也要避免模型 預測出階層上下雖有依賴關係,但其兩者都預測錯誤,造成嚴重的誤判性。所以 在量測階層式準確率時,應使用特定的指標來評估此模型的表現,而不是個別計 算各階層的 Top-1 準確率。準此,本論文提出三個階層式指標(綜合正確率、階 層關係一致性、預測風險率)來評估兩者階層式模型的預測好壞及預測錯誤時的 風險嚴重程度。在網路架構部分,本論文提出一個新的階層式架構模型及其訓練方法 HCN(Hierarchy-constrained Network),具體而言,該階層式模型在進行訓練時, 會將粗分類的特徵層融合到細分類的特徵層,同時也將模型的目標函數添加兩種 限制,第一種是限制粗細分類的特徵層,假設兩組影像同屬同一組粗分類,則這 兩張圖的這些特徵層數值要接近; 第二種則是模型預測出的粗分類及細分類類 別輸出需要保有一致關係(e.g. 細分類的狗就要對應到粗分類的動物而不是卡 車)。透過此架構與目標函數的設計,並測試在三種資料集上 (CIFAR100\\TinyImageNet-200\\CUB200-2011),我們發現此架構除了在階層式指 標有優良的表現外,因其訓練方式會使得同屬粗類別的細類別特徵層彼此數值更 相近,模型在受到對抗例攻擊時會較難將輸入影像的細分類判斷成與此影像不同 粗分類的細分類,以此來抵抗對抗例攻擊。 zh_TW dc.description.abstract (摘要) Hierarchical relationships exist in many datasets that involve object classification tasks. For example, animal recognition involves species classification, which is organized hierarchically into family, genus, and species. Similarly, ship classification in satellite images has finer sub-classes under main categories such as civilian and military vessels, including fishing boats, yachts, patrol ships, and destroyers. Therefore, it is crucial to design models capable of predicting hierarchical classification results.In hierarchical classification tasks, predicted answers should not only maintain hierarchical consistency (i.e., the predicted sub-class should be a child class of the predicted parent class), but also avoid serious misjudgments caused by incorrect predictions of both parent and child classes. Hence, it is important to design appropriate metrics to evaluate the performance of hierarchical models beyond calculating Top-1 accuracy for individual classes in each hierarchy. To address this issue, three new metrics, namely Aggregated Accuracy, Hierarchy Consistency, and Risk Factor, are proposed to accurately assess the performance of hierarchical models and the severity of prediction errors.In terms of network architecture, this thesis introduces a novel hierarchical model and training method called HCN (Hierarchy-constrained Network). During training, the features of the parent class are fused into the features of the child class. Additionally, two constraints are added to the objective function of the model. The first constraint considers the similarity of features between parent and child classes. If two sets of images belong to the same parent class, their features should be similar. The second constraint ensures that the predicted outputs of parent and child class categories maintain consistent relationships, such as the dog category of the child class corresponding to the animal category of the parent class, and not a truck category.Experimental analysis on three datasets, namely CIFAR100, TinyImageNet-200, and CUB200-2011, demonstrates that the proposed HCN outperforms existing hierarchical models based on the proposed hierarchical metrics. Furthermore, the features of child classes belonging to the same parent class are found to be more similar, making it difficult for the model to misclassify the sub-class of an input image into a sub-class of a different parent when subjected to adversarial attacks, thereby enhancing the overall robustness of the model. en_US dc.description.tableofcontents 摘要 iAbstract ii目錄 iii表次 v圖次 vii式次 viii第一章 緒論 11.1 研究背景與動機 11.2 研究目的與貢獻 11.3 論文架構 5第二章 技術背景與相關研究 6第三章 研究方法 113.1 階層式資料集和階層式評估指標 113.1.1 階層式資料集– CIFAR100\\ TinyImageNet200\\CUB 200-2011 113.1.2 階層式分類任務評估方法 143.2 階層式模型架構及其訓練方法 163.2.1 HCN階層式模型架構 163.2.2 HCN損失函數 17第四章 實驗結果 194.1 參數設定 204.2 CIFAR100\\TinyImageNet200多樣物件之實驗結果與討論 214.2.1 CIFAR100資料集上階層預測結果 214.2.2 TinyImageNet200資料集上階層預測結果 254.3 CIFAR100\\TinyImageNet200多樣物種之對抗例攻擊比較 294.3.1 對抗例攻擊介紹 294.3.2 對抗例實驗結果 304.4 消融實驗 334.4.1 HCN模型只由CE訓練以及架構改變帶來的影響 344.4.2 HCN模型少了ArcFace損失函數的影響 344.4.3 分群數量的變化所帶來的影響 364.4.3.1 CIFAR100 以K-Means++進行粗細分群 364.4.3.2 CIFAR100 分群數量減少 374.4.3.3 TinyImageNet200 分群數量增加 384.4.3.4 粗細分類按照隨機分群 394.5 三階層分類 404.5.1 三階層資料 (CIFAR100/ TinyImageNet200) 404.5.2 HCN模型三階層架構 414.5.3 三階層實驗結果 (CIFAR100/ TinyImageNet200) 424.5.4 三階層對抗例實驗結果 (CIFAR100/ TinyImageNet200) 454.6 CUB 200-2011單一物種的階層實驗結果與對抗例實驗結果 474.6.1 CUB 200-2011階層實驗結果 474.6.2 CUB 200-2011階層對抗例實驗結果 53第五章 結論與未來工作 55附錄 56參考文獻 61 zh_TW dc.format.extent 20306565 bytes - dc.format.mimetype application/pdf - dc.source.uri (資料來源) http://thesis.lib.nccu.edu.tw/record/#G0110753112 en_US dc.subject (關鍵詞) 深度學習 zh_TW dc.subject (關鍵詞) 神經網路 zh_TW dc.subject (關鍵詞) 階層式分類 zh_TW dc.subject (關鍵詞) 對抗例攻擊 zh_TW dc.subject (關鍵詞) Deep learning en_US dc.subject (關鍵詞) Neural network en_US dc.subject (關鍵詞) Hierarchical classification en_US dc.subject (關鍵詞) Adversarial attack en_US dc.title (題名) 階層式深度神經網路及其應用 zh_TW dc.title (題名) Deep Hierarchical Neural Networks and Its Applications en_US dc.type (資料類型) thesis en_US dc.relation.reference (參考文獻) [ 1 ] Roy, D., Panda, P., & Roy, K. Tree-CNN: a hierarchical deep convolutional neural network for incremental learning. Neural Networks, 121, 148-160. 2020.[ 2 ] Jiang, S., Xu, T., Guo, J., & Zhang, J. Tree-CNN: from generalization to specialization. EURASIP Journal on Wireless Communications and Networking, 2018(1), 216. 2018.[ 3 ] Xinqi Zhu, Michael Bain. B-CNN: Branch Convolutional Neural Network for Hierarchical Classification. arXiv 2017, arXiv:1709.09890.2017.[ 4 ] Salma Taoufiq , Balázs Nagy , Csaba Benedek.HierarchyNet: Hierarchical CNN- Based Urban Building Classification. Remote Sensing,Volume 12 ,Issue 22 .2020.[ 5] Yan, Z., Zhang, H., Piramuthu, R., Jagadeesh, V., DeCoste, D., Di, W., & Yu, Y.. HD-CNN: hierarchical deep convolutional neural networks for large scale visual recognition. In Proceedings of the IEEE international conference on computer vision (pp. 2740-2748). 2015.[ 6 ] M. D. Zeiler and R. Fergus. Visualizing and Understanding Convolutional Networks, pages 818–833. Springer Interna- tional Publishing, Cham, 2014.[ 7 ] A. Krizhevsky and G. Hinton. Learning multiple layers of features from tiny images. Computer Science Department, University of Toronto, Tech. Rep, 2009.[ 8 ] Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., & Fei-Fei, L. Imagenet: A large- scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition (pp. 248–255).2009.[ 9 ] Tan, M. and Le, Q.V. EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. Proceedings of the 36th International Conference on Machine Learning, ICML 2019, Long Beach, 9-15 June 2019, 6105-6114, 2019.[ 10 ] D. Arthur and S. Vassilvitskii. k-means++: The advantages of careful seeding. In SODA, pages 1027–1035, 2007.[ 11 ] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The caltech-ucsd birds-200-2011 dataset. 2011.[ 12 ] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In CVPR, 2016.[ 13 ] Deng J, Guo J, Xue N, Zafeiriou SJAPA .Arcface: Additive angular margin loss for deep face recognition, 2018.[ 14 ] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In International Conference on Learning Representations (ICLR), 2015.[ 15 ] Francesco Croce, Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks, 2020. zh_TW