Aiming at the continually increasing number of spams, an approach for spam filtering based on the use of Stacked Denoising Autoencoder (SDA) was proposed. Firstly, to get more abstract and robust feature representation of raw data, greedy layer-wise unsupervised algorithm was used to train the SDA by minimizing the construction error on unlabeled data set. Then a classifier was added on the top level of SDA. Next, the parameters of SDA were optimized with supervised algorithm by minimizing the classification error to obtain a optimal model on labeled data set. Lastly, experiments were performed on six different public corpora using the trained SDA. The performance of SDA algorithm was compared with Support Vector Machine (SVM), Bayes approach and Deep Belief Network (DBN), by using precision, recall, Matthews Correlation Coefficient (MCC) with more balanced performance measure as the experimental measures. The experimental results indicate that using SDA to filter spams has higher precision and more robustness. Since it not only acquires best average performance with all precision greater than 95%, but also gets close to prefect prediction with all MCC greater than 0.88.