A hybrid stream cipher scheme is proposed based on the novel interacting neural networks and the multiple chaotic systems. At first, random sequences generated by 3 independent logistics functions respectively are taken as dynamic inputs to 3 hidden layers of the interacting neural networks model. Then two inner weights of the two structures of neural networks will be synchronized through some steps of interacting learning, and the random key stream can be finally identified by combining the random sequence extracted from the aforementioned synchronized weight and 3 Logistics sequences. The comparison shows that the generated key stream performs the better randomness than others. As a good example, the proposed novel chaos-based stream cipher works perfectly on digital image encryption.