Many researchers have already shown that only user-based or content-based features are not enough to detect rumor in social media and for better prediction we need to consider both. In our research, we argue that the word embedding feature and sentiment score with subjectivity can also play a vital role in this detection task. Moreover, to detect the rumor at a very early stage and debunk it we may need to make the detection framework portable to legitimate users. This critical situation demands a secure implementation of rumor detection framework so that the user information used for training the prediction model can be protected from unauthorized access. In our experiment, we have also found that besides SVM, Logistic Regression and Random Forest algorithms, Artificial Neural Network and k-Nearest Neighbor can be used for rumor detection purpose where Artificial Neural Network and Random Forest outperformed (more than 90%) among all these algorithms in terms of accuracy. Other three algorithms also performed well with 80% or more accuracy level. To establish the robustness and efficiency of our proposed rumor detection mechanism, Precision, Recall, F1 Score, 10-fold Cross Validation, MCC, Confusion Matrix performance measures are used.