https://ir.lib.ncu.edu.tw/handle/987654321/91 Search the Channel s https://ir.lib.ncu.edu.tw/simple-search https://ir.lib.ncu.edu.tw/handle/987654321/110868 title: 音樂生成平台之實作與架構優化 Implementation and Architecture Optimization of a Music Generation Platform;音樂生成平台之實作與架構優化 Implementation and Architecture Optimization of a Music Generation Platform abstract: 隨著生成式AI的快速發展，音樂生成技術已取得重大突破，能夠從文字直接生成具備人聲與伴奏的完整歌曲。然而，現有的開源模型多針對英語或華語開發，且推論過程需要極龐大的運算資源，對普及化與本土語言的支持造成限制。本論文旨在於硬體資源受限的環境下（如限制於 8GB VRAM 的 NVIDIA RTX 3050），實作並優化一個高效能的音樂生成平台。 本研究以開源的 YuE 7B 基礎模型為核心，透過動態攔截注入（Monkey Patch） 技術，成功套用4-bit Normal Float 4 （4-bit NF4）量化，將模型顯示卡記憶體占用由原生的 28GB 大幅壓縮至約 7.7GB，有效避免記憶體不足錯誤。文中廣泛整理 HeartMuLa、LeVo 等最新音樂生成領域之技術發展，並透過實際建置 SongGen 與 YuE 模型進行推論對比，驗證單階段與雙階段自迴歸模型在生成穩定度與音質表現上之差異。 實驗結果顯示，優化後的系統能成功在消費級顯示卡上完成歌詞的輸入與音樂生成，並解決 Windows 環境下的編碼相容性錯誤及音訊儲存的碼率損失，提出具體的改良策略。此外，本平台亦奠定後續訓練與擴展至台語等本土語言生成的技術基礎。;With the rapid development of generative AI， music generation technology has made significant breakthroughs， enabling the direct generation of full songs with both vocals and accompaniment from text. However， existing open-source models are primarily developed for English or Mandarin， and their inference process requires massive computational resources， which limits their widespread adoption and support for local languages. This thesis aims to implement and optimize a high-performance music generation platform in hardware-constrained environments (such as being limited to an NVIDIA RTX 3050 with 8GB VRAM). Using the open-source YuE 7B foundation model as the core， this study successfully applies 4-bit Normal Float 4 (4-bit NF4) quantization through dynamic interception and injection via Monkey Patch technology. This significantly compresses the model′s GPU memory footprint from its native 28GB to approximately 7.7GB， effectively preventing out-of-memory errors. The paper comprehensively reviews recent technological advancements in the music generation field， such as HeartMuLa and LeVo. Furthermore， through the practical implementation and inference comparison of the SongGen and YuE models， it verifies the differences in generation stability and audio quality between single-stage and two-stage autoregressive models. Experimental results demonstrate that the optimized system can successfully process lyrics input and music generation on consumer-grade GPUs. It also resolves encoding compatibility errors and audio storage bitrate loss in the Windows environment， proposing specific improvement strategies. Furthermore， this platform lays the technical foundation for future training and expansion into the generation of local languages such as Taiwanese. <br> Tue, 20 Oct 2026 08:33:01 GMT https://ir.lib.ncu.edu.tw/handle/987654321/110865 title: StegAV1:基於幀內編碼預測角度修改與係數嵌入之AV1視訊資料隱藏;StegAV1:基於幀內編碼預測角度修改與係數嵌入之AV1視訊資料隱藏 abstract: 本研究探討AV1視訊編碼資料隱藏，改良基於調整幀內編碼之預測角 度嵌入方法，並分析係數嵌入方法的影響。既有的AV1Stego方法主要於 位元流輸出階段修改預測角度資訊，證明於AV1中進行隱寫術的可行 性。然而，該嵌入結果不再回到前端預測與殘差計算流程，於是在高嵌入 量條件下常造成畫質劣化。本研究將資料嵌入時機調整至率失真最佳化完 成後、最終模式資訊寫入前的階段，降低資料於候選模式更新過程中被覆 蓋的可能，同時也減少對額外控制參數的依賴。 StegAV1 於全幀內（Intra-only）與關鍵幀（Key frame）畫面的方向性 預測區塊進行資料嵌入，以奇偶對映方式調整角度偏移值，解碼端依據角 度偏移值平移後的奇偶性還原隱藏位元。StegAV1在不同位元率條件下皆 能穩定完成資料嵌入與擷取，且在中高位元率編碼下維持良好的客觀畫 質，在嵌入穩定性與視覺品質維持間取得良好平衡。此外，於AV1中以 係數修改的方法嵌入資訊，資訊隱藏量與畫質表現會受到若干編碼條件所 影響，本研究建議依據不同情況調整嵌入條件，降低資料擷取不同步的風 險與畫質損耗，可作為後續更高容量AV1視訊資料隱藏設計的參考。;This study investigates data hiding in AV1 video coding by improving a prediction-angle-based embedding method that modifies intra-frame coding and analyzes the effects of coefficient-based embedding. Existing AV1Stego primarily modified prediction-angle information at the bitstream output stage， demonstrating the feasibility of steganography in AV1. However， the embedded information of AV1Stego is not fed back into the preceding prediction and residual calculation processes so high embedding payloads can result in noticeable visual degradation. In this study， the embedding process is moved to the stage after rate-distortion optimization and before the final mode information is written， thereby reducing the possibility that the embedded data will be overwritten during candidate-mode updates while minimizing the reliance on additional control parameters. StegAV1 embeds data into directional prediction blocks in both intra-only and key frames. The hidden bits are embedded by adjusting the parity of angular offsets， while the decoder recovers the hidden bits according to the parity of the angular offset after the corresponding shift. Experimental results show that StegAV1 can reliably embed and extract data under different bitrate conditions. It also maintains good objective visual quality at medium-to-high bitrates， achieving a favorable balance between embedding reliability and visual quality preservation. In addition， when coefficient modification is adopted for data embedding in AV1， both the embedding capacity and visual quality are affected by several coding conditions. We therefore recommend adjusting the embedding conditions according to different coding scenarios to reduce the risk of ii synchronization failures during data extraction and minimize visual quality degradation. These findings can serve as a reference for the development of higher-capacity data hiding schemes for AV1 video. <br> Tue, 20 Oct 2026 08:32:43 GMT https://ir.lib.ncu.edu.tw/handle/987654321/110862 title: 基於市場微觀結構與熱力學特徵之台灣加權股價指數高頻交易預測與交易策略研究;A Study on High-Frequency Trading Prediction and Trading Strategies for the Taiwan Stock Exchange Capitalization Weighted Stock Index (TAIEX) Based on Market Microstructure and Thermodynamic Features abstract: 金融市場價格變動受到市場供需、投資人行為及資訊流動等因素影響，傳統投資分析多依賴產業面、基本面、財務面及技術面資訊進行決策，然而短期價格波動往往無法完全由上述因素解釋。近年來，高頻交易與人工智慧（Artificial Intelligence， AI）技術快速發展，使市場微觀結構（Market Microstructure）資訊逐漸成為金融市場研究的重要方向。市場微觀結構能反映價格形成過程中的供需變化，可能提供較價格本身更早的市場訊號。此外，金融市場具有高度複雜性與非線性特徵，與熱力學系統存在相似之動態行為，因此本研究進一步導入熱力學概念量化市場狀態。 本研究以臺灣證券交易所（Taiwan Stock Exchange Corporation， TWSE）提供之台灣加權股價指數（Taiwan Stock Exchange Capitalization Weighted Stock Index， TAIEX）5秒高頻資料為研究對象，利用委託買賣資訊、成交資訊及買賣失衡指標建立市場微觀結構特徵，並建構市場能量（Market Energy）、市場熵（Market Entropy）及市場溫度（Market Temperature）等熱力學特徵。研究中分別採用極端梯度提升（eXtreme Gradient Boosting， XGBoost）、長短期記憶網路（Long Short-Term Memory， LSTM）與轉換器（Transformer）三種AI模型進行未來60秒價格方向預測，並提出熱力學風險覆蓋層（Thermodynamic Risk Overlay， TRO），透過市場熵與市場溫度辨識高風險市場環境，以降低交易風險。 實驗結果顯示，市場微觀結構資訊具備高頻價格預測能力，而導入TRO後，各模型之累積報酬率與風險控制能力皆獲得改善。進一步透過 Walk-Forward Expanding Window Validation 驗證，結果顯示本研究方法在不同市場環境下均具有良好穩健性與泛化能力。 綜合研究結果可知，市場微觀結構與熱力學特徵能有效提升交易策略績效，所提出之TRO亦能改善風險管理能力。本研究提供一套結合市場微觀結構、AI模型與風險控制機制之高頻交易研究架構，可作為未來量化交易與金融科技研究之參考。;Financial market price movements are influenced by various factors， including market supply and demand， investor behavior， and information flow. Traditional investment analysis primarily relies on industry trends， fundamental analysis， financial indicators， and technical analysis for decision-making. However， short-term price fluctuations cannot always be fully explained by these factors. With the rapid development of high-frequency trading (HFT) and artificial intelligence (AI)， market microstructure information has become an increasingly important research area in financial markets. Market microstructure reflects the supply–demand dynamics during the price formation process and may provide earlier market signals than price information itself. Furthermore， financial markets exhibit highly complex and nonlinear characteristics that resemble the dynamic behavior of thermodynamic systems. Therefore， this study introduces thermodynamic concepts to quantify market conditions. This study uses 5-second high-frequency data of the Taiwan Stock Exchange Capitalization Weighted Stock Index (TAIEX) provided by the Taiwan Stock Exchange Corporation (TWSE). Market microstructure features are constructed from order flow information， transaction information， and order imbalance indicators. In addition， thermodynamic features， including Market Energy， Market Entropy， and Market Temperature， are developed to characterize market conditions. Three artificial intelligence models， namely eXtreme Gradient Boosting (XGBoost)， Long Short-Term Memory (LSTM)， and Transformer， are employed to predict future price movements over the next 60 seconds. Furthermore， a Thermodynamic Risk Overlay (TRO) is proposed to identify high-risk market environments through market entropy and market temperature， thereby reducing trading risk. Experimental results demonstrate that market microstructure information possesses predictive power for high-frequency price movements. After incorporating the Thermodynamic Risk Overlay， all models achieve improvements in cumulative returns and risk control performance. Furthermore， Walk-Forward Expanding Window Validation is conducted to evaluate the robustness and generalization ability of the proposed approach. The results indicate that the proposed method maintains stable performance across different market conditions. Overall， the findings suggest that market microstructure and thermodynamic features can effectively enhance trading strategy performance， while the proposed Thermodynamic Risk Overlay improves risk management capability. This study provides a high-frequency trading framework that integrates market microstructure， artificial intelligence models， and risk management mechanisms， offering valuable insights for future quantitative trading and financial technology research. <br> Tue, 20 Oct 2026 08:32:31 GMT https://ir.lib.ncu.edu.tw/handle/987654321/110859 title: TEE-Industry：基於 ARM TrustZone 之即時與安全工業控制系統架構;TEE-Industry: A TrustZone-Based Architecture for Real-Time and Secure Industrial Control Systems abstract: 在工業 4.0 的浪潮下，現代工業控制系統（Industrial Control Systems， ICS）日益部署於通用運算平台之上，因而衍生新的安全性與即時效能挑戰。在此類環境中，關鍵控制程序必須與複雜且非關鍵的應用程式共用同一套硬體平台，使控制邏輯暴露於潛在的干擾與網路攻擊之中。 本論文提出 TEE-Industry，一套運用 ARM TrustZone 技術之可信執行環境（Trusted Execution Environment， TEE）保護工業控制迴路的系統架構。與先前將控制迴路置於 OP-TEE（S-EL1）之 TEE-PLC 設計不同，TEE-Industry 將控制迴路及其時間關鍵周邊置於 ARM Trusted Firmware-A (TF-A) 所在的 EL3，藉此消除每個週期中 S-EL1 與 EL3 之間的世界切換，並縮小可信運算基礎（Trusted Computing Base， TCB）。 在此可信執行環境之內，即時排程機制與安全周邊存取控制策略共同確保控制迴路準時執行，並維持其感測器與致動器輸入輸出的可信性，即使主作業系統遭到攻陷亦然。 我們於 Rockchip RK3399 平台上進行初步評估，量測到每個週期的固定軟體開銷約為 1.0 毫秒，約占標稱 100 毫秒控制週期的 1%；其往返時間的變異亦較對應的一般世界基準更小。上述結果為工業 4.0 時代的安全且可靠的工業控制，提供一個深具前景的解決方案。 關鍵字：ARM TrustZone、可信執行環境、工業控制系統、可程式邏輯控制器、即時系統、Modbus RTU、安全周邊存取;Modern industrial control systems (ICS) under the Industry 4.0 paradigm increasingly run on general-purpose computing platforms， which introduces new security and real-time performance challenges. In these environments， critical control processes must operate alongside complex， non-critical applications on a shared hardware platform， exposing the control logic to potential interference and cyber-attacks. This thesis proposes TEE-Industry， an architecture that leverages a Trusted Execution Environment (TEE) on ARM TrustZone technology to protect ICS control loops. Unlike prior TEE-PLC designs that host the control loop at S-EL1 inside OP-TEE， TEE-Industry places the loop and its time-critical peripherals at EL3 inside ARM Trusted Firmware-A (TF-A)， eliminating the S-EL1↔EL3 context switch on each cycle and shrinking the trusted code base. Within the TEE， a real-time scheduling mechanism and a secure peripheral access control strategy keep the control loop timely and its sensor/actuator I/O trustworthy， even if the main operating system is compromised. Our initial evaluation on a Rockchip RK3399 platform measures a fixed software overhead of approximately 1.0 ms per cycle—about 1 % of the nominal 100 ms control period—and shows tighter round-trip variance than an equivalent Normal World baseline， providing a promising solution for secure and reliable industrial control in the era of Industry 4.0. Keywords: ARM TrustZone， Trusted Execution Environment， Industrial Control Systems， PLC， Real-Time Systems， Modbus RTU， Secure Peripheral Access <br> Tue, 20 Oct 2026 08:32:27 GMT