We propose a diffusion expectation-maximization algorithm with adaptive combiner for distributed estimation over sensor networks. Due to the spatial distribution of the nodes, variation of node profile across the network is a common phenomena in real applications. The unreliable nodes exist and provide inaccurate estimates, which may be caused by high levels of noise or malicious attacks. Instead of using a static combiner, an efficient adaptive combination scheme is developed by formulating it as a ℓ0-norm regularized minimum variance unbiased estimation problem. The proposed algorithm is robust to the variation of node profile across the network. Furthermore, it can be extended for node-specific processing. Each node estimates a subset of the global variable of the whole network.