A large number of people are suffering from diabetes worldwide and the number of patients is increasing at an alarming rate. Conventionally, invasive methods are employed for the detection of diagnosis. Here, we propose a non-invasive method that can distinguish between diabetic and non-diabetic patients accurately and efficiently. The method involves Pulse Plethysmograph (PuPG) sensor to obtain bio-signals from the human body and utilization of different machine learning (ML) classifiers such as support vector machine (SVM), K-Nearest Neighbor (KNN), fine and bagged tree for diagnosis of diabetes. Our dataset includes 276 signals from individuals out of which 100 are diabetic and 176 are non-diabetic. First, we normalized the obtained signals and then removed the unwanted noise from each signal by applying empirical mode decomposition (EMD). Secondly, we extracted 12 different salient features from the noise-free signals. Finally, we obtain the classification accuracy of our approach by using 4 different classifiers. Fine-KNN gives the maximum accuracy of 94.0%