Abstract: From the requirement of robust classification capability of binary feedfor-ward neural networks (BNNs), the notion of robust classification is provided, and the general equation form of robust classification hyperplanes is presented. From this general equation, if a BNN is constructed by only robust neurons, each connection weight can take only -1, 0 or+1 and the threshold of each neuron in the net is a simple addition of base threshold of 1/2 and a complement shreshold ranged in a valid integer region. Also presented in this paper is that the threshold of each neuron is a summation of contributions from each input of this neuron. All of these characteristcs are of much importance in constructing a BNN with the greatest robustness, less number of hidden neurons and less number of connections in the net. It is also shown that the robust BNN particularly facilitates VLSI implementation.