邁向強健的網路安全：聯邦學習與網路入侵偵測之規避攻擊研究

Abstract

隨著 5G 通訊技術普及以及大量物聯網終端接入網路，現代網路逐步演變為分散且組成多元的系統型態，使得整體運作架構更加繁複，也同步放大了潛在資全威脅。在此情境下，現存的系統於實際部署時逐漸面臨限制，包括在流量環境中難以有效擴充、需分析的網路流量資料量急速膨脹，以及對於新興且持續變化的攻擊為缺乏調適能力。 近幾年來，物聯網成為發動資安攻擊的重要載體，相關事件不僅數量快速增加，惡意軟體在功能設計與攻擊策略上也更加多樣且精細。這類感染活動已上升為全球性的安全風險，原因是一旦應變處理延誤，會造成嚴重損害，如敏感資訊被非法存取或外洩。現有的防護與事件回應措施仍不足以有效因應此類威脅，從而凸顯出強化即時偵測機制，以及建立跨利害關係者之間高效率資訊共享體系的需求。 本論文以防禦端與攻擊端的雙重視角切入網路安全問題。首先，提出 FINISH 的分散式協同學習框架，彙整來自多個暗網感測節點的觀測資訊，在無須集中原始資料同時能建構模型與分析。此方法不僅能顯著降低節點之間的通訊成本，也能達到隱私保護需求，並具備應用於 5G 多接取邊緣運算場域的可行性。此外，論文亦提出並深入探討一種新穎的規避技術──流量偽裝攻擊，其核心概念在於將惡意行為巧妙地融合進看似正常的流量，從而有效規避現有資安系統的識別能力。 藉由將聯邦學習理念納入資安體系，FINISH 即使面對來源異質且分散的資料環境，仍能展現一致且可信的入侵偵測能力。相較之下，FCA 在極低的計算成本下即可實現高度有效的規避效果，從而暴露出當前資安系統於架構與假設上的核心缺陷。綜合來看，結果強調了在對抗的威脅中，重新評估協作式資安機制之必要性，並為後續建構兼具韌性、可擴充性與隱私保護能力的資安方案奠定了關鍵基礎。

The swift growth of modern networking environments, driven by the rollout of 5G systems and the explosive growth of Internet of Things (IoT) devices, has resulted in significantly more intricate infrastructures and a much broader attack surface. Consequently, traditional Network Intrusion Detection Systems (NIDS) are increasingly strained by issues such as poor scalability, massive traffic volumes, and the necessity to respond to fast-evolving threat behaviors. In recent years, large-scale cyberattacks leveraging IoT devices have become alarmingly frequent, with IoT malware growing in both diversity and sophistication. These infections pose significant global risks, as delayed incident responses often result in severe consequences such as the leakage of sensitive information. Current countermeasures remain insufficient, underscoring the urgent need for early detection and effective information sharing among stakeholders. To address these challenges, this dissertation proposes innovative defense and attack mechanisms that (1) operate effectively in decentralized and large-scale environments and (2) preserve data privacy, and (3) advanced evasion techniques. This work tackles the problem from both defensive and offensive perspectives. First, it introduces FINISH, a federated learning framework that integrates decentralized darknet sensor data using nonnegative matrix factorization. FINISH reduces communication overhead, supports privacy-preserving collaboration, and scales to 5G multiaccess edge computing scenarios. Second, it presents the Flow Camouflage Attack (FCA), a novel evasion strategy that conceals malicious activity within benign-like network flows, effectively bypassing state-of-the-art NIDS. By integrating federated learning into network defense, FINISH demonstrates robust detection capabilities across heterogeneous data sources. In parallel, FCA achieves complete evasion with minimal computational cost, exposing critical vulnerabilities in modern NIDS. Together, these findings emphasize the necessity of evaluating collaborative intrusion detection systems under adversarial pressure and pave the way for resilient, scalable, and privacy-preserving network security solutions.