Model Uncertainty Based Annotation Error Fixing for Web Attack Detection

Xinyu Gong, Jialiang Lu, Yuefu Zhou, Han Qiu, Ruan He

Abstract: Deep learning (DL) techniques have been widely used in web attack detection domain. With the stronger ability to fit data, DL models are also more sensitive to the training data, annotation errors can mislead the model training very easily. In our work, we propose model uncertainty to evaluate the prediction made by the DL based web attack model. Model uncertainty helps to find the annotation errors and also the misclassification caused by these errors. Experiments on two real web log datasets and a public benchmark dataset prove that our method can find more annotation errors than the traditional DL method. We also prove the efficiency of our detection model: with the aid of GPU acceleration, it is capable of tackling attacks on the fly.