Fentanyl and morphine are opioid drugs as well as new psychoactive substances. Even originally introduced as efficient anesthetic drugs to relieve moderate-to-severe pain in clinic, the overdose of new synthetic opioids is currently a serious public health problem in numerous countries worldwide. The entire category of fentanyls has been included in the regulatory list in several countries. There is a great and urgent demand to rapidly recognize fentanyls and morphines in various samples. Here, we report an on-site surface-enhanced Raman spectroscopic method to classify fentanyls from morphines by the Raman spectroscopic signature of the molecular scaffold structure, with an assistance of principle component analysis algorithm. Moreover, by simple but fine-tuning approach of inorganic salt-induced aggregation of gold nanoparticles substrate, we achieved a selective detection of 10 ng/mL fentanyl from 2000-fold of heroin, the most common coexisting substance in chemical samples. Good differentiation of 50 ng/mL fentanyl from 10 000-fold morphine as a main metabolite of heroin in urine samples was also possible after a feasible pretreatment by StageTip procedures. Depending on different structures, the detection sensitivity of five fentanyls ranged from 50 to 2000 ng/mL.