In network analysis, the salient network structure of interest, instead of being directly observed, is often hidden in a larger network in which most structures are not informative. The noises introduced by such overwhelming yet non-informative data can obscure the salient structure and limit the effectiveness of many network analysis methods. Traditionally, this scenario is modeled as a core-periphery structure, and algorithms are designed to extract the core. Subsequent analyses are then focused on the core, which is believed to have more pronounced structural patterns. In this paper, we propose a network model for non-informative structure of networks without imposing specific form for the core. Specially, we assume that the non-informative nodes connect to other nodes in a purely random pattern (up to individual node degree), while the core structure can take any informative pattern. Moreover, an algorithm of core extraction is proposed by leveraging spectral properties of the network. The algorithm is computationally efficient and has a theoretical guarantee of accuracy. The effectiveness of the algorithm is demonstrated through extensive simulations and real data examples.