In this paper, a novel microscopic modeling strategy is proposed to investigate the effective thermal conductivity of composites with consideration of stochastic interface defects. To this end, the subdomain boundary element method combined with asymptotic homogenization is proposed to effectively solve the thermal conduction problem. In order to accurately capture the heat flux on the boundary and the internal region in the representative volume element (RVE), a parameterized sub-cell is constructed to discretize the RVE. On this basis, the influence of stochastic interface defects on the thermal conductivity of composites is investigated by utilizing the Monte Carlo method. Specifically, the effect of the location, length, thickness, and area of the interface defects on the thermal conductivity is investigated. A proportional decrease in the transverse thermal conductivity coefficient is found for interface defect areas ranging from 1% to 10%.