The concept of cognitive reserve (CR) originated from discrepancies between the degree of brain pathology and the severity of clinical manifestations. CR has been characterized through CR proxies, such as education and occupation complexity; however, such approaches have inherent limitations. Although several methods have been developed to overcome these limitations, they fail to reflect the entire Alzheimer's disease (AD) pathology. Meanwhile, graph theory analysis, one of most powerful and flexible approaches, have established remarkable network properties of the brain. The functional and structural brain networks are damaged in neurodegenerative diseases. Therefore, network analysis has been applied to clarify the characteristics of the disease or give insight. Here, using multimodal neuroimaging, we propose an intuitive model to estimate CR based on its original definition, and explore the neural substrates of CR from the perspective of networks and functional connectivity. A total of 87 subjects (21 AD, 32 mild cognitive impairment, and 34 normal aging) underwent tau and amyloid PET, 3D T1-weighted MR, and resting-state fMRI. We hypothesized CR as a residual of actual cognitive performance and expected performance to be related to quantitative factors, such as AD pathology, demographics, and a genetic factor. Then, we correlated this marker using education and occupation complexity as conventional CR proxies. We validated this marker by testing whether it would modulate the effect of brain pathology on memory function. To examine the neural substrates associated with CR, we performed graph analysis to investigate the association between the CR marker and network measures at different granularities in total subjects, AD spectrum and normal aging, respectively. The CR marker from our model was well associated with education and occupation complexity. More directly, the CR marker was revealed to modify the relationship between brain pathology and memory function among AD spectrum. The CR marker was correlated with the global efficiency of the entire network, nodal clustering coefficient, and local efficiency of the right middle-temporal pole. In connectivity analysis, one cluster of edges centered on right middle-temporal pole was significantly correlated with the CR marker. In subgroup analysis, the network measures of right middle-temporal pole still correlated with the CR marker among AD spectrum. However, right precentral gyrus was revealed to be associated with the CR marker in normal aging. This study demonstrates that our intuitive model using multimodal neuroimaging and network perspective adequately and comprehensively captures CR. From a network perspective, CR is associated with the capacity to process information efficiently in the brain. The right middle-temporal pole was revealed to be a pivotal neural substrate of CR in AD spectrum. These findings foster understanding of AD and will be useful to help identify individuals with vulnerability or resistance to AD pathology, and characterize patients for intervention or drug trials.