基於光電流洩漏之旁通道攻擊：在有限資料集中的特徵選取到模型評估

Abstract

本研究聚焦於光電流模擬資料中之潛在資訊洩漏，提出一套模組化且可擴展的旁通道攻擊分析流程。透過模擬加密晶片在執行進階加密標準-128（Advanced Encryption Standard-128, AES-128）運算時的光子行為與電流變化，我們獲得一組具時程整合特性的資料集，每筆觀測之旁通道揭露跡 （side-channel trace, 以下簡稱揭露跡）皆包含 62,724 個由三種物理通道（Iavg、Irms、Ipeak）所構成的特徵值。實驗重點針對不同的特徵選取策略（白箱、黑箱、灰箱）以不同的物理通道來源和攻擊模型（統計式與機器學習式）進行交叉分析與比較。研究發現，透過適當的標準化處理與互資訊（mutual information, MI）選取法，三通道聯集（union）特徵來源能有效整合各通道之互補資訊，達到本研究中最佳的攻擊效能。我們並透過特徵重疊程度分析與晶片空間上的特徵點分布觀察，揭示了統計式選取法與晶片先驗知識之間的特徵選取差異。本研究進一步嘗試透過標籤空間之轉換（漢明權重降維與子位元拆解降維）來因應標籤不平衡（imbalanced）與資料量有限的問題。並引入猜測熵（guessing entropy）、猜測熵曲線下面積（area under the guessing entropy curve, AUGE）以及正規化後的猜測熵曲線下面積（normalized area under the guessing entropy curve, N-AUGE）來當作旁通道攻擊無法完美收斂時的跨攻擊模型之成效評估指標。實驗結果顯示，灰箱攻擊特徵選取方法（相關係數選取法與互資訊選取法）能超過白箱攻擊已知興趣點（known point of interest, known POI）特徵之效能；同時，漢明權重降維與子位元拆解降維的應用亦能提升旁通道攻擊模型的效能與穩定性。此研究不僅驗證了光電流模擬資料可作為旁通道攻擊之研究基礎，也為未來特徵選取與標籤空間的設計提供具體方向與策略建議。

This thesis investigates the potential information leakage embedded within photonic current simulation data and proposes a modular, extensible framework for side-channel analysis. By simulating the photonic behavior of an Advanced Encryption Standard-128 (AES-128) encryption chip during operation, we obtain a dataset comprising 62,724 features per trace, constructed from three statistical channels: average (Iavg), root mean squared (Irms), and peak current (Ipeak). We evaluate multiple attack models, including traditional statistical approaches: correlation power analysis (CPA), mutual information analysis (MIA), and machine learning classifiers: k-nearest neighbor (KNN), Gaussian naïve Bayes (GNB), support vector machine (SVM), random forest (RF), extreme gradient boosting (XGBoost), across various feature selection strategies: white-box (known points of interest, known POI), black-box (variance-based), and gray-box (correlation-based and mutual information-based). Furthermore, we explore label space reduction techniques—such as Hamming weight (HW) transformation and bit-splitting—to address challenges posed by the label imbalanced and limited dataset. Experimental results show that gray-box feature selection strategies based on correlation and mutual information outperform the white-box attacks. Additionally, label reduction techniques enhance model stability and classification performance. This study confirms the utility of simulated photonic side-channel data in security analysis and offers practical strategies for feature selection and label reduction design in the data-constrained scenarios.