RS Attack Detection
Old School Method (Before 2016)
2016年之前Shilling Attack Detection大多是从人工设计特征,制定几个统计数据,从而区别正常用户和恶意用户。手工做好的特征输入到传统ML分类器如决策树等中,举两个例子比如:
Preventing shilling attacks in online recommender systems (WIDM’05)
设计两个特征DegSim和RDMA:
Degree of similarity with Top Neighbors:
Degsim u = ∑ v = 1 k W u , v k \operatorname{Degsim}_{u}=\frac{\sum_{v=1}^{k} W_{u, v}}{k} Degsimu=k∑v=1kWu,v
W u , v W_{u, v} Wu,v是用户u和v的相似度。所以这个指标是用户u和最相似的k个user的平均相似度。
Rating Deviation from Mean Agreement (RDMA):
R D M A u = ∑ i = 0 N u ∣ r u , i − r ˉ i ∣ N R i N u R D M A_{u}=\frac{\sum_{i=0}^{N_{u}} \frac{\left|r_{u, i}-\bar{r}_{i}\right|}{N R_{i}}}{N_{u}} RDMAu=Nu∑i=