Intent of malicious users to use social networks for malicious and inappropriate purposes has encouraged them to create fake accounts or compromise legitimate accounts. Latter is a crucial aspect as it involves risk to a user’s reputation as well as confidential data. Though researchers have proposed different techniques to deal with the detection of compromised accounts but they have analyzed it from a single perspective be it text, metadata, network, or time series. In this paper, a multidimensional user profiling-based technique has been proposed that models features from different dimensions and employs authorship verification procedure to detect the compromised accounts in online social networks. The task is to utilize a multidimensional feature set vector and assess which features and algorithm as well as what set of fine-tuned parameters help achieve better results. Both binary (two-class) and unary (one-class) classification perspectives have been taken into consideration for the detection of compromised accounts. For binary classification scenario, all the considered classifiers, namely, k-NN, RF, GB, SVM, and MLP attained an F-score above 96%. (SVM with rbf kernel outperformed others obtaining an average F-score of 97.38%). Likewise in Unary Classification, OCC-SVM (rbf kernel) outperformed other three unary classifiers (LOF, IF, and OCC-SVM) achieving an average F-score of 89.64% and Matthews correlation coefficient of 76.41%.