We propose a novel approach to attenuate multiple reflections using randomized-order empirical mode decomposition (EMD). EMD is applied in each frequency slice to obtain different oscillating wavenumber components. The several highest wavenumber components are then removed to remove the unflattened high wavenumber components. In order to make EMD based smoothing approach capable of attenuating spatially coherent noise, we propose to first apply a trace randomization procedure to the NMO-corrected CMP gather to make the unflattened multiple reflections spatially incoherent (like random noise) and thus easier for attenuation. Compared with the other denoising approaches that are based on spatial coherency, the proposed EMD based approach is non-parametric and can obtain a better demultiple performance with a better preservation of the flattened primary reflections. The proposed EMD based approach is also more adaptive than those more advanced EMD based approaches, like EEMD Wu and Huang (2009), CEEMD Colominas et al. (2012), or ICEEMD Colominas et al. (2014), which require more input parameters.