Counterfeit electronics form a major roadblock towards a safe and successful economy. An increase in globalization has led to a major increase in the total number of counterfeit products all around the world. While several methods have been designed to detect counterfeits, very few of them have been applied to the system-on-chip (SoC). The influx of a variety of components in SoCs and the conglomeration of different types of properties makes it difficult to detect counterfeit SoCs. In this paper, we aim at detecting recycled counterfeit SoCs by evaluating the degradation of power supply rejection ratio (PSRR) of a low drop-out (LDO) regulator, a principal component of the power supply of the SoC. Since the power supply is a universal component in all SoCs, this method can be considered effective for most SoCs. We apply machine learning (ML) algorithms pertaining to the family of Gaussian mixture models to classify SoCs as recycled or new. Supervised and unsupervised ML algorithms show an accuracy of up to 90% and 74% of recycled detection. We also apply stand-alone LDO PSRR degradation to train the ML algorithm and test on PSRR from embedded LDOs in SoCs. This form of semi-supervised ML performed well for our previous experiments of recycled detection with stand-alone LDOs but was not able to distinguish recycled SoCs from new SoCs, thus increasing the number of false detection.