| 摘要: | 推薦系統的使用隨著電子商務的急速成長而被廣泛地運用在各行各業，推薦系統可以協助使用者在複雜的環境中做決策，其主要是根據推薦系統本身對使用者和推薦的項目的瞭解，進而推薦項目給使用者，藉此幫助消費者快速找到他們可能會感興趣的資料或產品，其中最常被推薦系統採用的推薦方法為協同過濾式，然而，由於推薦系統處在一個開放式的環境裡，很容易被惡意的攻擊者在評分資料中注入假資料，也就是製造不實的評分、意見或評論，企圖推升或打壓某些項目。
近幾年來，許多偵測攻擊的演算法相繼被提出，然而這些方法的應用仍然受到不同的限制，PCA在偵測攻擊的相關論文有很好的成果，但其缺點在於無法處理在使用者項目矩陣中的遺漏值，導致整個計算出來的結果有很大的誤差，另外，利用SPC的統計方法在做資料分群時，需要每個項目所累積的長期資料，才能作有意義的分群進而訓練數據。
本研究以Beta分配的特性設計了一個偵測攻擊的方法，此方法沒有遺漏值的問題，資料分群也不需要累積一定時間的大量資料，就可以偵測出被攻擊的產品，另外，有別於過去的學者皆假設攻擊者在使用者資料中只攻擊一個目標項目，本研究在實驗中模擬攻擊者攻擊多個目標項目，Beta分配的方法其偵測率達到70%以上，錯殺率則介在10%到20%內。
;With the rapid growth of e-commerce, Recommender system is widely used in various walks of life. Recommender systems can help users make decisions in a complex environment, which is mainly based on the understanding of user and item by Recommender system, and then recommend the item to the user. Thereby helping consumers quickly to find the information or products they may be interested in. The most commonly recommended method is Feature-based Filtering. However, due to the recommender system is in an open type of environment, malicious attacker is easy to inject false information into the rating data, which means manufacture false ratings, opinions or comments to push up or suppress certain items.
In recent years, many detection algorithms have been proposed, however the application of these methods are still subject to different restrictions. For instance, the related papers of the PCA-based method have good results, but its drawback is that the PCA-based cannot handle missing values, leading to a great error. In addition, the use of statistical methods in doing SPC (Statistical Process Control) data clustering, you need a long-term cumulative data for each item in order to make a meaningful grouping and thus the training data.
In this study, we made use of the characteristics of the Beta distribution method to design attack detection. This method can solve the problem of missing values, besides, data clustering does not need to accumulate large amounts of data for some time and it can detect the products which is attacked. In addition, there are other in the past scholars are assuming that the attacker only attacks one target item in the user data, this study simulated lots of attackers attack multiple targets in the experiment. Beta distribution method for the Detection rate is over 70%, the False alarm rate is between 10% to 20%. |
