In intrusion detection based on traditional BP networks, the model of the BP network algorithm is easy to fall into local optimum and the initial value is random. So how to choose initial value directly affects the training effect of BP networks. We therefore propose an improved harmony search (HS) algorithm to optimize the initial value of BP neural networks, which gets the BP neural network a set of better initial values. Experimental results show that the improved HS algorithm has higher fitness function value, and the detection rate and convergence rate of the algorithm are improved when the BP network optimized by this algorithm is used in intrusion detection.