類別不平衡下敵對式無線感知模型之標註與學習

Abstract

WiFi 通道狀態資訊（CSI）感測技術的持續發展引發了重大的隱私疑慮，因為竊聽者可能透過牆壁監控室內的人類活動。儘管現有研究已證實穿透牆壁的 WiFi 感測技術的可行性，但仍存在一個關鍵限制：大多數研究假設可以取得標記的訓練數據，這在攻擊者無法直接觀察目標活動的對抗性場景中是不切實際的。為了解決這一挑戰，我們提出了 WiSneak 框架，能夠在不需直接觀察目標的情況下實現基於 CSI 的敵對式活動識別。WiSneak 包含兩個關鍵階段：動作標記階段，結合時間模式分析與機器學習模型，從側通道資訊中自動生成訓練標籤；以及動作識別階段，整合自監督預訓練與數據處理，有效處理訓練標籤中的類別不平衡問題。通過全面的實驗，我們的動作標記階段在「躺下」和「站起」動作檢測中分別達到了 0.94 和 0.91 的 F1 分數。在包含五種人類活動的真實世界 CSI 數據集上評估時，我們的動作識別階段展現了優異的性能，在高度不平衡的條件下（少數類別「跑步」和「跌倒」分別只有 5 個和 2 個標記樣本，而多數類別「行走」有 330 個標記樣本），F1 分數達到 0.69，比現有的半監督方法提升了 36.2%。該框架的有效性不僅揭示了 WiFi 感測技術對隱私的潛在威脅，也透過強調現實世界敵對式感測系統的實際限制與能力，為設計更有效的防禦機制提供了寶貴的見解。

The continuous advancement of WiFi Channel State Information (CSI) sensing technology poses significant privacy concerns, as eavesdropper can potentially monitor indoor human activities through walls. While existing research demonstrates the feasibility of through-wall WiFi sensing, a critical limitation remains: most studies assume access to labeled training data, which is unrealistic in adversarial scenarios where attackers cannot directly observe the target's activities. To address this challenge, we propose WiSneak, a novel framework that enables adversarial CSI-based activity recognition without requiring direct observation of the target. WiSneak consists of two key stages: a Motion Labeling Stage that combines temporal pattern analysis with a custom-designed model to automatically generate training labels from side-channel information, and a Motion Recognition Stage that integrates self-supervised pretraining with adaptive data processing to effectively handle class imbalance in the generated labels. Through comprehensive experiments, our Motion Labeling Stage achieves 0.94 and 0.91 F1-scores for "lay down" and "stand up" motion detection respectively. When evaluated on a real-world CSI dataset containing five human activities, our Motion Recognition Stage demonstrates superior performance with a macro F1-score of 0.69, outperforming existing semi-supervised approaches by 36.2% under highly imbalanced conditions where minority classes "run" and "fall" have only 5 and 2 labeled samples respectively, compared to 330 labeled samples for the majority class "walk". The proposed framework's effectiveness not only reveals concerning implications for WiFi sensing privacy but also provides valuable insights for designing more effective defense mechanisms by highlighting the practical constraints and capabilities of real-world adversarial sensing systems.