Feature selection, structure design and weight training are considered as three key tasks for the application of neural network. Traditional leaning algorithms of neural network, which only optimize one or two aspects of these three tasks, neglect the fact that these three tasks are interdependent and make a united contribution to the performance of neural network. In order to model normal behaviors accurately and improve the performance of intrusion detection, a joint evolutionary neural network (JENN) is presented in this paper. Input features, network structure and connection weights are evolved jointly using genetic algorithm. To evolve all three tasks jointly, a hybrid representation scheme is employed, and penalty factors for the number of input nodes and hidden nodes are introduced into fitness function. The generated-subnet based crossover operator is adopted in consideration of the relationship between genotype and phenotype. Experimental results with the KDD-99 dataset show that the proposed JENN achieves better detection performance in terms of detection rate and false positive rate.