| 摘要: | 隨著智慧行動裝置的普及，如何為計算和通訊能力有限的邊緣或區域裝置研發具有隱私保護功能的分散式學習演算法及資料融合技術已成為建立更彈性通訊網路架構的重要研究議題。分散式壓縮學習策略結合了壓縮感測技術以有效減少目標訊號還原所需的量測數目，並利用感測矩陣作為保密金鑰進行資料加密。這類合作學習方法目前已被廣泛應用於當前的人工智慧通訊領域。然而，現有的合作學習方法多建立於能夠同時獲得取樣後樣本的振幅與相位資訊的基礎條件下進行。本論文是第一篇探討分散式壓縮相位檢索之課題，即研究在僅有限量測振幅資訊的情況下，如何讓網路中的區域裝置透過合作學習一起完成稀疏訊號的重建任務。考慮到通訊網路中的區域裝置可能受地理位置或感測能力限制的影響，不同於現有文獻大都是基於完整視角的訊號模型，即單一全域訊號，進行演算法開發，本研究針對只能觀測到部分景象資訊的用戶裝置群，研發高效能且具隱私保護的合作與訊號壓縮方法，以共同實現對全域訊號及區域訊號的恢復。另外，有鑑於次世代網路通訊系統需要實現高傳輸率與低延遲性的服務品質，我們讓網路中的區域裝置採用稀疏感測(Sparsesensing)技術來有效率地獲取量測數據。我們首先利用目標訊號與感測矩陣的稀疏結構特性設計出低運算複雜度的合作學習演算法，完成全域訊號的非零位置估計。接下來基於此估計結果並透過解最小方差問題得到全域訊號非零值的最佳估計。最後基於全域訊號估計結果，利用交替乘子演算法完成區域訊號還原。值得注意的是，透過上述的分散式合作學習策略，網路系統中的區域裝置端可以在不分享各自的量測資料的情況下，合作完成各自的訊號重建任務。因此，我們提出的還原方法具有一定的隱私保護功能。另外，為了減少在執行 交替乘子演算法過程中所需的運算成本及手動調整最佳參數的不便，本研究進一步採用深度展開技術提出模型驅動的新型深度學習架構，旨在提升整體合作學習的能源效益。模擬結果顯示：即便網路中各ii個區域裝置所獲得的量測數量遠低於獨立還原訊號所需的量測數量，我們所提出的合作學習方法仍然可以在不需分享量測原始資料的情境下，透過彼此之間的合作完成各自的區域訊號還原；其還原效能比現行許多專為壓縮相位檢索設計的重建方法有更高的準確度。;etd.lib.nycu.edu.tw/cgi-bin/gs32/ncugsweb.cgi/ccd=PziWQO/login?jstimes=1&loadingjs=1&userid=guest&o=dwebmge&cache=1722348367122With the widespread adoption of smart mobile devices, developing distributed learningalgorithms and data fusion techniques with privacy protection features for low-cost edge orlocal devices has become a crucial research topic. Distributed compressive learning integratescompressive sensing techniques to significantly reduce the measurements required for sparsesignal reconstruction, and also utilize the sensing matrix as a cryptographic key for dataencryption. Such collaborative learning methods are now widely applied in the field of AIcommunications. However, existing collaborative learning methods mainly rely on thecondition that both amplitude and phase information can be simultaneously obtained from themeasurements during the sampling process. This paper is the first to address the issue ofdistributed compressive phase retrieval, specifically investigating how local devices within anetwork can collaboratively reconstruct sparse signals with only limited measurementamplitude information.Considering the potential limitations due to geographical location or sensing capabilitiesof local devices in networks, this study examines a multi-view compressive phase retrievalproblem where each device can observe only a partial view of the global sparse signal. Here,partial view means that certain arbitrary and unknown indices of the global signal vector areunobservable to that device and, therefore, do not influence the measurement outcomes. In thiswork, the objective for each local device is to reconstruct its local partial-view signal vectorusing its own measurements and messages received from the central server about the globalsignal, all while preserving the privacy of each device. In addition, recognizing the need forhigh transmission rates and low latency in next-generation network communication systems, aivsparse sensing technique is employed at each local device to reduce data gathering and storagecosts. By exploiting the sparse nature of both the signal and sensing matrices, the global signalsupport can be identified using a simple counting rule. Then based on this support estimate, thecorresponding nonzero signal entries in the global signal vector can be recovered by solving aleast squares problem. Subsequently, a distributed alternating direction method of multipliers(ADMM)-based iterative algorithm is proposed for local sparse signal reconstruction. Notably,our proposed collaborative distributed learning scheme enables joint reconstruction of acommon sparse signal by local clients without the need to share local datasets with the centralserver, while also recovering personalized sparse vectors, i.e., partial-view vectors.Consequently, the proposed recovery method incorporates notable privacy protection features.In order to reduce the computational costs and the inconvenience of manually tuning theoptimal parameters during the execution of the ADMM-based algorithm, this study furtheradopts deep unfolding techniques to propose a new model-driven deep learning architectureaimed at enhancing the energy efficiency of collaborative learning. Simulation results show thateven though the number of measurements obtained by local devices in the network is far lessthan what is needed for independent signal reconstruction, our proposed collaborativedistributed learning algorithm can still achieve stable signal reconstruction through cooperationwithout sharing raw measurement data. Furthermore, the estimation accuracy of our methodsignificantly outperforms that of many existing methods specifically designed for compressivephase retrieval. |
