Click frauds are Internet crimes where clicks are deliberately performed to increase the publisher's earnings or to deplete an advertising budget of the advertiser's competitor. This paper presents an approach to automatically detecting click frauds by using a mathematical theory of evidence to estimate the likelihood of the frauds from click behaviors. Unlike most existing work, our approach provides online computation that incorporates incoming behaviors on real-time. It is theoretical-grounded, easily extensible and employs only data available at the advertiser's site. The latter makes the approach feasible for the advertisers to guard themselves against the frauds. The paper describes the approach and evaluates its validity using real-world case scenarios.