Hydraulic property of fractured rock masses is commonly undertaken based on 2-D fracture network models, which are cut planes of the real 3-D models. This simplification would lead to a significant underestimation of fracture network permeability. In this study, a numerical procedure is originally developed to address flow problem through 3-D discrete fracture network (DFN) models. In this method, fractures are modeled as circular discs with arbitrary size, orientation and location. Fracture networks are established with fractures following well-known statistical distributions, after which the networks are triangulated and fluid flow is calculated by solving the Reynolds equation using Galerkin method. The results show that the permeability of 2-D DFN models that are cut from an original 3-D DFN model would be underestimated by 19.2 ∼ 43.6%, comparing with that of the 3-D DFN model. For networks that are consisted of power-law size-distributed fractures, the equivalent permeability would decrease exponentially with the increasing length distribution exponent. This tendency can be interpreted by incorporating the average intersection length, which is a parameter that can reflect the connectivity of a fracture network. When the heterogeneity of fracture aperture distribution is considered, some tortuous flow paths are formatted in 3-D fracture networks. The rougher fracture surface, the stronger anisotropy of aperture distribution, and thereby resulting in the larger reduction of the network permeability.