Yes this is possible. Although quantification requires careful labeling to reduce unspecific binding. Labeling degree also has to be considered. Multiple detection of single molecules can be reduced by an optimized lighting scheme or careful analysis if labeling density is low enough.

In principle, yes. However, you have to ensure two things if you want to work quantitatively: First, you need an exactly defined amount of label on the protein, i.e. exactly one, two, three, otherwise the signal will correspond not to the amount of protein (maybe you could think about incorporation of unnatural amino acids and, for example, click chemistry, but this can be tricky). Second, you need to make sure that you have no unspecific labeling of other proteins. And of course, the detection must be at the single-molecule level;)

Detection of single molecules is the foundation of super-resolution localization microscopy (STORM/PALM).  These techniques are often called "single molecule localization microscopy," or SMLM.

Detecting a molecule below the diffraction limit simply requires that its signal be isolated spatially or temporally from other molecules in a sample.  If that signal is very strong, sampled multiple times - or both - then its position in each dimension can be localized with high precision.  Insightful techniques that exploit this approach to break the diffraction limit garnered a Nobel Prize in 2014.

While detection is not difficult in principle, quantification is trickier.  It is theoretically possible for photoactivated localization microscopy (PALM) because one can control the ratio of label to target (e.g. tagging a single fluorescent protein to one's protein of interest).  So a single localization should correspond to a single molecule.  However, as noted by Marc Renz, the photophysical properties of your reporter (a dye or fluorescent protein) need to be carefully characterized to avoid counting errors.  For example, the detection efficiency and photoblinking characteristics of fluorescent proteins dictate how they are counted.

Labeling density and the multimeric nature of protein complexes also influences the counting strategy.  There are several computational approaches to untangling densely labeled structures/clusters, including pair correlation and dark time/blinking analysis. These approaches generally rely on on modeling and probability distributions to extract information from patterns of individual molecule localizations.

Could anyone please tell me how STORM detects single molecule? What is the main principle of STORM (Stochastic Optically Reconstruction Microscope)?