http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/197 2026-08-19T10:45:21Z http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/102440 標題: 高效光譜嵌入演算法：理論、樣本複雜度與應用研究; Efficient Spectral Embedding Algorithms: Theory, Sample Complexity, and Applications 作者: 葉行遠; Sing-Yuan Yeh 摘要: 光譜嵌入 (Spectral embedding) 方法是流形學習 (Manifold learning) 與非線性降維中極為重要的工具，能夠透過圖拉普拉斯算子的光譜分析，揭示高維資料內在的幾何結構。本論文從理論與演算法兩個層面出發，對光譜方法進行改進，提出兩種以代表性樣本 (landmark) 為基礎的新型演算法，以提升其計算效率與可擴展性。第一種方法為 Landmark Alternating Diffusion（LAD），加速了傳統應用於感測器融合任務的 Alternating Diffusion（AD）架構。我們在流形假設下為其提供理論保證，並於多通道腦波（EEG）之睡眠階段標註任務中驗證其效能。第二種方法為 Landmark Vector Diffusion（LAVENDER），為 Vector Diffusion Maps（VDM）於大規模資料上的可擴展版本，能有效捕捉如影像區塊模型等複雜資料結構，並具備一致性與收斂性的理論保證。為了提升彈性與穩健性，這兩個方法皆引入使用者可調整的正規化參數，用以修正取樣分布並減少對代表性樣本選擇的敏感性。綜上所述，本論文提出具理論基礎的新型光譜嵌入演算法，顯著提升其在實際資料分析任務中的可擴展性與穩定性。; Spectral embedding methods are fundamental tools in manifold learning and nonlinear dimensionality reduction, enabling the recovery of intrinsic data geometry through the spectral analysis of graph Laplacians. This thesis advances both the theoretical and algorithmic aspects of spectral methods by introducing two landmark-based variants designed for scalability and improved efficiency. The first method, Landmark Alternating Diffusion (LAD), accelerates the vanilla Alternating Diffusion (AD) framework used in sensor fusion. We provide theoretical guarantees under manifold assumptions and demonstrate its effectiveness in EEG-based sleep stage annotation. The second method, Landmark Vector Diffusion (LAVENDER), a scalable version of vanilla Vector Diffusion Maps (VDM) designed to capture complex data structures, such as those in patch-based image modeling. We also establish theoretical results for LAVENDER, including consistency and convergence guarantees. To enhance flexibility and robustness, both methods incorporate user-defined normalization parameters to correct sampling density and mitigate sensitivity to landmark selection. Overall, this thesis contributes novel, theoretically grounded algorithms that enhance the scalability and reliability of spectral embedding methods in practical data analysis tasks. 2025-01-01T00:00:00Z http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/84701 標題: 零樣本歌聲轉換與合成的統一模型; A Unified Model for Zero-Shot Singing Voice Conversion and Synthesis 作者: Jui-Te Wu; 吳睿得 摘要: 深度學習的最新進展不僅促進了零樣本歌聲合成和歌聲轉換任務的 實現，同時也提供了將這兩個任務統一為一個通用模型的機會。在本 文中我們提出了一個統一兩項任務的模型，可以從文本或音頻格式的 任意源歌唱內容生成任意目標歌手的歌聲。該模型結合了處理文本輸 入的詞源編碼器以及處理音頻輸入的聲源編碼器進行訓練，並透過以 動態規劃為基礎的自督導式學習，編碼器將會在訓練過程中學習如何 將音頻與音素進行最佳的對齊。這些編碼器也將音頻和文本數據分別 映射到一個相似的潛在空間中，使得歌聲轉換與合成兩項任務可以透 過同一個解碼器來完成。目標歌手的參考音檔被轉換成以幀為單位的 碎片化資訊，並透過注意機制來根據源內容進行提取與重構，這使模 型能夠在測試階段從文本或音頻源生成沒學習過的目標歌手的聲音。 客觀和主觀實驗都證實，所提出的模型表現超越過去最佳的任意歌聲 轉換與任意歌聲合成模型。; Recent advances in deep learning not only facilitate the implementation of zero-shot singing voice synthesis (SVS) and singing voice conversion (SVC) tasks, but also provide the opportunity to unify these two tasks into one gen- eralized model. In this paper, we propose such a model that can generate singing voice of any target singer from any source singing content in either text or audio format. The model incorporates self-supervised joint training of the phonetic source encoder and the acoustic source encoder, with an audio- to-phoneme alignment process in each training step, such that these encoders map the audio and text data respectively into a shared, temporally aligned, and singer-agnostic latent space. The target singer’s latent representations en- coded at different granularity levels are all trained to match the source latent representations sequentially with the attention mechanisms in the decoding stage. This enables the model to generate unseen target singer’s voice with fine-grained resolution from either text or audio sources during the inference stage. Both objective and subjective experiments confirmed that the proposed model is competitive with the state-of-the-art SVC and SVS methods. 2022-01-01T00:00:00Z http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/101689 標題: 隱私保護神經網路推論之優化方法; Optimization Methods for Privacy-Preserving Neural Network Inference 作者: 顧昱得; Yu-Te Ku 摘要: 本論文旨在提升隱私保護的深度神經網路（DNN）推論之可部署性，聚焦於解決主要的效率瓶頸，也就是非線性運算（例如 ReLU），同時維持可量化的安全性與隱私保證。第一條研究主線聚焦於第三代全同態加密（FHE），特別是 FHEW/TFHE。為了降低加密推論成本，我提出 FHE-Neuron，透過重新配置FHEW/TFHE 參數與 bootstrapping 的資料流，並採用動態密文精度切換，以取得更佳的準確度與延遲折衷。我也進一步發展一套 FHE-aware 的量化與微調框架，使預訓練模型能更適應加密執行環境，並減輕由精度限制與噪聲所引入的誤差。第二條研究主線提出 DP Protocol，這是一種受差分隱私啟發的協定安全性放寬定義，要求在相鄰輸入下，對手所觀察到的視圖分佈必須保持接近，從而能在可調整的隱私界限下達到更高效率。基於此概念，我設計一個伺服器輔助的兩方安全計算（2PC）推論協定：線性層由兩方以秘密分享方式計算；非線性運算則在「隱私增強」資料上以明文執行，而這些資料由加噪與秘密打亂所產生。我並以形式化方式證明該協定滿足 DP-Protocol 的保證。總結而言，本論文從加密執行與協定設計兩個面向推進實務可用的安全推論，並展示共同設計密碼機制、模型架構，以及誤差與隱私分析，能夠顯著提升隱私保護 DNN 推論的效率與可用性。; This dissertation improves the deployability of privacy-preserving deep neural network (DNN) inference by addressing the dominant efficiency bottleneck in nonlinear computation (e.g., ReLU), while maintaining quantifiable security and privacy guarantees. The first line of work focuses on third-generation Fully Homomorphic Encryption (FHE), particularly FHEW/TFHE. To reduce the cost of encrypted inference, I propose FHE-Neuron, which reconfigures FHEW/TFHE parameters and the bootstrapping dataflow, and employs dynamic ciphertext precision switching to achieve improved accuracy–latency trade-offs. I further develop an FHE-aware quantization and fine-tuning framework to adapt pretrained models to encrypted execution and mitigate precision- and noise-induced errors. The second line of work introduces DP-Protocol, a Differential Privacy–inspired relaxation of protocol security that requires an adversary’s view distributions to remain close for neighboring inputs, enabling higher efficiency under tunable privacy bounds. Building on this notion, I design a server-aided two-party computation (2PC) inference protocol in which linear layers are evaluated via two-party secret sharing, while nonlinearities are executed in plaintext on “privacy-enhanced” data produced by noise injection and secret shuffling. I formally prove that the protocol satisfies the DP-Protocol guarantee. Overall, this dissertation advances practical secure inference from both encrypted execution and protocol design perspectives, demonstrating that co-designing cryptographic mechanisms, model architecture, and error and privacy analysis can substantially improve the efficiency and usability of privacy-preserving DNN inference. 2026-01-01T00:00:00Z http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/85054 標題: 針對記憶體受限裝置執行模式匹配的正規表達式分群演算法; Evolutionary Rule Balancer: A RegEx Algorithm for Memory-Constrained Devices to Perform Pattern Matching 作者: Pang-Cheng Wu; 吳邦誠 摘要: 模式匹配是各種重要應用程序的關鍵，例如在封包檢測，它偵測惡意封包以保護網絡。為了使應用程序更強大，我們可能需要在資源受限的設備上執行模式匹配。例如，嘗試在智能網卡上應用DPI，以使我們的服務能夠在短時間內處理更多的封包。但是，DPI的內存需求可能比 NIC 所擁有的記憶體量大得多。因此，我們嘗試將正規表達式 分群，讓每個運算資源只需要構建較小的自動機。 為了實現這個想法，我們設計了一種通用的分群算法，稱為Evolutionary Rule Balancer。通過合適的估計和評估函數，該算法能夠適用於各種自動機模型。整個系統由FA Engine、Matched String Generator、FA Size Estimator和Intelligent Optimization Algorithm組成。該算法是有效率的，因為我們透過估計自動機大小避免構造過多的資動機。另一方面，利用 RegEx 的嵌入模型，我們的做法也支持動態更新。根據實驗結果，對於在多個 NFA 上執行的模式匹配，Rule Balancer 可以獲得更好的結果。; Pattern matching is the key component of various important applications such as deep packet inspection, which detects malicious packets to protect our network. To make the applications be more stronger, we may need to perform pattern matching on resource-constrained devices. For instance, we try apply DPI on smart network interface card to make our service be able to handle more packets in a short time. However, the memory space requirement can be much larger than NIC have. Therefore, we try to divide RegExes into several groups and build smaller FA for each core. To realize the idea, we design a general grouping algorithm, called Evolutionary Rule Balancer. With suitable estimation and evaluation functions, the algorithm should be able to work for various FA models. The entire system consists of FA Engine, Matched String Generator, FA Size Estimator and Intelligent Optimization Algorithm. This method is efficient as we estimate FA size to avoid too many FA constructions and the entire algorithm supports dynamic update by utilizing RegEx Embedding Model. Our experiments show that for pattern matching performed on multi-NFAs, Rule Balancer achieves better results. 2022-01-01T00:00:00Z