Real data may present a significant amount of noise, generated by inaccuracies in data collection, transmission and storage. The presence of noisy data in a training dataset used for the induction of a Machine Learning model may increase the training time and the complexity of the induced model, resulting in the deterioration of its predictive performance for new data. Noise may be found in the input and target attributes. In this study, we are concerned with noise in the class label of the target attribute. For such, we propose and experimentally investigate some simple class noise detection and elimination strategies for classification problems, introducing controlled noise levels in five UCI datasets originally free of inconsistencies. The results obtained in the experiments performed show the potential of the proposed approaches.