Maybe machine learning techniques for translating text could produce unstripped binaries based on stripped binaries, given a large training set?