Collaborative Filtering, though a successful recommendation technique is vulnerable to shilling attacks due to its open nature. These attacks alter recommendations being generated for the user by inserting fake user profiles in the database. To minimize the bias introduced in the recommendation process, many machine learning methods have been explored and shown excellent results. However, supervised machine learning detection techniques are restricted to hand-designed features while unsupervised detection techniques require prior knowledge about fake profiles. In this paper, we propose a novel approach namely, ShillDetector for the detection of shilling attacks based on the recently proposed swarm intelligence technique, grey wolf optimization. The proposed approach works as a dimensionality reduction technique taking advantage of high correlation among shillers and removing correlated features that are redundant. Further, it works directly on the rating matrix, does not require hand-designed features, prior knowledge of attack profiles, or any training time. The performance of ShillDetector has been evaluated on the MovieLens dataset consisting of 100 K ratings. Experimental results depict that ShillDetector outperformed two state-of-the-art approaches, namely, SVM-TIA and PCA-VarSelect approaches with an average precision of 0.99 in case of average attack taken over different attack sizes, viz, 1%, 2%, 5%, and 10%.

协同过滤，尽管成功的推荐技术由于其开放性而容易受到先令攻击。这些攻击通过在数据库中插入虚假用户配置文件来改变为用户生成的推荐。为了尽量减少推荐过程中引入的偏差，已经探索了许多机器学习方法并显示出优异的结果。然而，有监督的机器学习检测技术仅限于手工设计的特征，而无监督的检测技术需要关于假配置文件的先验知识。在本文中，我们提出了一种新方法，即基于最近提出的群智能技术灰狼优化来检测先令攻击的 ShillDetector。所提出的方法作为一种降维技术，利用 shiller 之间的高相关性并去除冗余的相关特征。此外，它直接在评级矩阵上工作，不需要手工设计的特征、攻击概况的先验知识或任何训练时间。ShillDetector 的性能已经在包含 10 万个评级的 MovieLens 数据集上进行了评估。实验结果表明，ShillDetector 优于两种最先进的方法，即 SVM-TIA 和 PCA-VarSelect 方法，在平均攻击范围不同的情况下，即 1%、2% 的平均精度为 0.99 、5% 和 10%。攻击概况或任何训练时间的先验知识。ShillDetector 的性能已经在包含 10 万个评级的 MovieLens 数据集上进行了评估。实验结果表明 ShillDetector 优于两种最先进的方法，即 SVM-TIA 和 PCA-VarSelect 方法，在平均攻击范围不同的情况下，即 1%、2% 的平均精度为 0.99 、5% 和 10%。攻击概况或任何训练时间的先验知识。ShillDetector 的性能已经在包含 10 万个评级的 MovieLens 数据集上进行了评估。实验结果表明 ShillDetector 优于两种最先进的方法，即 SVM-TIA 和 PCA-VarSelect 方法，在平均攻击范围不同的情况下，即 1%、2% 的平均精度为 0.99 、5% 和 10%。