隨著網路使用的頻率逐漸增加，日常生活中越來越多的的行為開始與網路掛勾，當網路發生異常情況時，將會對我們的生活影響劇烈。然而，造成網路異常的原因非常的多樣，如病毒入侵、硬件故障、網路攻擊等。在這篇論文中，我們將聚焦在網路流量方面的異常進行偵測，以及早發現問題並進行異常的排除。我們提出了一種集成了空間與時間的分析的模型(Attentiongraph)方法。Attentiongraph 結合了 Attention GRU 處理空間方面的特徵與 GCN 處理時間方面的特徵以發現流量中的異常。在兩個真實世界的數據集上面的實驗結果表明，Attentiongraph 優於目前最先進的模型，達成了更高的準確性;As Internet services become more prevalent and ubiquitous today, our daily lives gradually depend on a stable and reliable Internet connection. Therefore, we will be prone to significant inconvenience when anomalous network events occur. To avoid this problem, it is crucial that we detect and resolve any network anomaly in a timely manner. In this thesis, we propose a system, named Attentiongraph, that utilizes both spatial and temporal features to detect network anomalies. To maximize the success of anomaly detection, Attentiongraph integrates the Attention mechanism, Attention GRU, and GCN into the system in the right order. To mitigate the data imbalance problem that is common in anomaly detection research, Negative Edge Selection is applied to ensure the model training stability. Experimental results on two real-world network datasets show that Attentiongraph is superior to current state-of-the-art deep learning anomaly detection models in terms of the AUC score.
