An important problem in causal inference is to break down the total effect of treatment into different causal pathways and quantify the causal effect in each pathway. Causal mediation analysis (CMA) is a formal statistical approach for identifying and estimating these causal effects. Central to CMA is the sequential ignorability assumption that implies all pre-treatment confounders are measured and they can capture different types of confounding, e.g., post-treatment confounders and hidden confounders. Typically unverifiable in observational studies, this assumption restrains both the coverage and practicality of conventional methods. This work, therefore, aims to circumvent the stringent assumption by following a causal graph with a unified confounder and its proxy variables. Our core contribution is an algorithm that combines deep latent-variable models and proxy strategy to jointly infer a unified surrogate confounder and estimate different causal effects in CMA from observed variables. Empirical evaluations using both synthetic and semi-synthetic datasets validate the effectiveness of the proposed method.