Machine learning models encounter plethora of challenges due to the changing data over time. This phenomenon is known as concept drift. Existing techniques for concept drift detection have shown promising results, but they require true labels as a precondition for drift detection to be successful. True labels are limited and expensive, especially in real-world application settings. To deal with this problem, this thesis proposes an AutoEncoder based Drift Detector (AEDD) technique for drift detection, that can detect drifts without access to true labels. This study combines two different techniques to achieve this. First, reconstruction error is measured by an autoencoder and then from the measured reconstruction error, using the ADaptive sliding WINdow (ADWIN) technique identifies structural changes over time. The observed drifts are utilized to retrain the prediction model. The thesis demonstrates the superiority of the technique by showing that the AEDD outperforms alternative state-of-the-art algorithms for classification tasks on real world datasets with artificially induced drift.