本研究提出了一種新穎的圖像檢索模型防禦機制，透過針對性的圖像轉換有效降低與後門攻擊相關的風險。利用如移除像素及水平翻轉等操作，並根據RISE技術生成的顯著圖動態調整策略，我們的方法破壞了嵌入圖像中的潛在觸發器。這些設置經過大量測試，以確保它們在不影響系統功能的情況下保持乾淨樣本的準確率。實驗結果表明，我們的防禦不僅優於傳統方法，還有效地對抗了先進的圖像檢索後門攻擊，大幅提升了圖像檢索系統的安全性。這種方法能夠讓圖像檢索系統運作不受影響，在正常操作條件下保持高精準度和功能性，並在不需要大規模重新訓練模型或更改系統設計的情況下有效地消除威脅。;This research introduces a novel defense mechanism for image retrieval models that effectively mitigates risks associated with backdoor attacks through targeted image transformations. By utilizing strategic techniques such as the removal of lines or columns of pixels and horizontal flipping, and dynamically adjusting transformations based on saliency maps generated by the RISE technique, our method disrupts potential triggers embedded within the images. These adaptations are refined through extensive testing to ensure they maintain the Mean Average Precision (MAP) of clean samples without adversely affecting system functionality. Experimental results demonstrate that our defense not only outperforms traditional methods but also effectively counteracts advanced image retrieval backdoor attacks, significantly enhancing the security of image retrieval systems. This approach allows the image retrieval system to operate efficiently, preserving high accuracy and functionality under normal operating conditions, and effectively neutralizing threats without extensive retraining or system redesign.
