SPDP: 用於減少日誌異常檢測分割推理階段中間通訊開銷的語意保留動態剪枝方法

Abstract

在系統日誌異常偵測情境中，近期研究多利用深度學習模型將大量人類語言的資料分析轉換並檢測。本研究提出一種「語意保留的 Token 剪枝方法」，以減少 Split Inference 架構中產生的通訊閘頓。我們所設計的模型可在無需標註資料的前提下，利用餘弦相似度將剪枝後與未剪枝的向量計算出 Loss，以此去學習出最適當剪枝比例的預測器，並搭配 L2 norm 作為 token 的重要性分數的作法。可以在不影響原先異常檢測模型的表現下，保留一定語意資訊，並盡可能地減少 activation 大小。實驗結果顯示，所提出方法可在不影響原先的準確率下，在多種日誌資料集上大幅降低 activation 傳輸量。

In practical scenarios of system log anomaly detection, deep learning models are commonly used to analyze and transform large volumes of data for abnormal behavior detection. This study proposes a semantics-preserving token pruning method to reduce the intermediate communication overhead during the Split Inference architectures. The proposed model requires no labeled data for training and learns a predictor that determines the optimal pruning ratio. By integrating L2-norm as token importance scores, our method effectively preserves semantic information while minimizing the size of activations, without degrading the performance of the original anomaly detection model. Experimental results show that the proposed method can significantly reduce activation transmission size across multiple log datasets, while maintaining comparable anomaly detection accuracy.