基於圖節點重要性之圖還原攻擊

Abstract

本文提出了一個針對圖逆推攻擊的方法。針對圖形神經網絡的安全和隱私問 題，提出了一個新的模型，在只需要訪問原始圖的節點嵌入矩陣，而無需與節點 嵌入模型進行交互，也不需要得知原始圖嵌入模型的算法，就可以實現還原原始 圖並獲得相當高的準確度。透過預測節點的連接數以及使用自編碼器來學習圖形 的結構，達到精準的預測原始圖形的結構。並經由實驗，展示了圖形還原攻擊的 有效性和實用性。

The thesis discusses the privacy risks associated with graph embedding models, particularly highlighting the possibility of a graph embedding inversion attack. It introduces a novel graph recovery attack capable of accurately reconstructing graph edges from node embeddings without any interaction with the embedding models or knowledge of the embedding algorithms. This is achieved by predicting node degrees and using an autoencoder to learn the graph’s properties, thus facilitating a precise reconstruction of the original graph. This raises significant privacy concerns. The effectiveness of this attack has been demonstrated through experimental validation.