Online auction fraud is becoming a serious problem in the recent years. To protect normal traders from fraud, effective countermeasures are needed to uncover cunning fraudsters. A lot of work has been proposed for this purpose. However, seldom of them took the detection cost into consideration, which is crucial for building a practical fraud detection system. To provide cost-effective fraud detection for online auction, a genetic feature selection method-CostGen is proposed in this paper, which can reduce the detection cost while improving the overall detection accuracy.