I believe that the BOLD time series is modeled according to various known confounds and the residuals are considered to be the 'true' noiseless signal.  Here is an excerpt from the original CONN paper:

"The toolbox implements an anatomical aCompCor strategy (Behzadi et al., 2007) in which a user- defined number of orthogonal time series are estimated using principal component analysis (PCA) of the multivariate BOLD signal within each of these noise ROIs. This strategy generalizes the common practice of extracting the average BOLD time series from one or several seeds located within the white matter and/or CSF areas. In addition, and for each original temporal confounding factor, first- and higher- order derivatives of the associated time series can also be de- fined by the user as additional confounding factors (e.g., Fox et al., 2005). Each of the defined temporal confounding factors is then regressed from the BOLD time series at each voxel (sep- arately for each session), and the resulting residual time series are band-pass filtered."

There's more in the paper if you want details.