Abstract: Internet users heavily rely on web search engines for their intended information. The major revenue of search engines is advertisements (or ads). However, the search advertising suffers from fraud. Fraudsters generate fake traffic which does not reach the intended audience, and increases the cost of the advertisers. Therefore, it is critical to detect fraud in web search. Previous studies solve this problem through fraudster detection (especially bots) by leveraging fraudsters' unique behaviors. However, they may fail to detect new means of fraud, such as crowdsourcing fraud, since crowd workers behave in part like normal users. To this end, this paper proposes an approach to detecting fraud in web search from the perspective of fraudulent keywords. We begin by using a unique dataset of 150 million web search logs to examine the discriminating features of fraudulent keywords. Specifically, we model the temporal correlation of fraudulent keywords as a graph, which reveals a very well-connected community structure. Next, we design DFW (detection of fraudulent keywords) that mines the temporal correlations between candidate fraudulent keywords and a given list of seeds. In particular, DFW leverages several refinements to filter out non-fraudulent keywords that co-occur with seeds occasionally. The evaluation using the search logs shows that DFW achieves high fraud detection precision (99%) and accuracy (93%). A further analysis reveals several typical temporal evolution patterns of fraudulent keywords and the co-existence of both bots and crowd workers as fraudsters for web search fraud.