mRR result of searching the original image.

Image search systems could be endangered by adversarial attacks and data perturbations. The image retrieval system can be compromised either by distorting the query or hacking the ranking system. However, existing literature primarily discusses attack methods, whereas the research on countermeasures to defend against such adversarial attacks is rare. As a defense mechanism against the intrusions, quality assessment can complement existing image retrieval systems. “GuaRD” is proposed as an end-to-end framework that uses the quality metric as a weighted-regularization term. Proper utilization and balance of the two features could lead to reliable and robust ranking; the original image is assigned a higher rank while the distorted image is assigned a relatively lower rank. Meanwhile, the primary goal of the image retrieval system is to prioritize searching the relevant images. Therefore, the use of leveraged features should not compromise the accuracy of the system. To evaluate the generality of the framework, we conducted three experiments on two image quality assessment(IQA) benchmarks (Waterloo and PieAPP). For the first two tests, GuaRD achieved enhanced performance than the existing model: the mean reciprocal rank(mRR) value of the original image predictions increased by 61%, and the predictions for the distorted input query decreased by 18%. The third experiment was conducted to analyze the mean average precision (mAP) score of the system to verify the accuracy of the retrieval system. The results indicated little deviation in performance between the tested methods, and the score was not effected or slightly decreased by 0.9% after the GuaRD was applied. Therefore, GuaRD is a novel and robust framework that can act as a defense mechanism for data distortions.

图像搜索系统可能面临对抗性攻击与数据扰动的威胁。图像检索系统可通过两种路径遭到破坏：一是篡改查询请求，二是入侵排序系统。然而，现有研究多聚焦于攻击方法本身，针对此类对抗性攻击的防御对策却相对匮乏。作为抵御入侵的防御机制，质量评估可对现有图像检索系统形成有效补充。本研究提出名为GuaRD的端到端框架，将质量指标作为加权正则化项融入其中。对两类特征进行合理利用与均衡配置，可实现可靠且鲁棒的排序：原始图像将被赋予更高的排序优先级，而经过篡改的图像则会获得相对较低的排序结果。与此同时，图像检索系统的核心目标是优先返回与查询相关的图像。因此，合理利用此类融合特征不应损害系统的检索精度。为验证该框架的通用性，研究团队在两个图像质量评估（Image Quality Assessment, IQA）基准数据集Waterloo与PieAPP上开展了三组对比实验。在前两组测试中，GuaRD的性能相较现有模型实现了显著提升：原始图像预测的平均倒数排名（Mean Reciprocal Rank, mRR）值提升了61%，针对篡改后查询输入的预测结果则下降了18%。第三组实验通过分析系统的平均精度均值（Mean Average Precision, mAP）得分，验证检索系统的检索精度。实验结果显示，各测试方法的性能偏差极小；引入GuaRD框架后，系统的mAP得分未受影响，仅出现0.9%的微小下降。综上，GuaRD是一种新颖且鲁棒的框架，可作为抵御数据篡改攻击的有效防御机制。