Locate-Then-Detect: Real-time Web Attack Detection via Attention-based Deep Neural Networks

Authors: Tianlong Liu, Yu Qi, Liang Shi, Jianan Yan

IJCAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments are carried out on both benchmarks and real Web traffic.
Researcher Affiliation Collaboration Tianlong Liu 1 , Yu Qi 2, , Liang Shi 3 and Jianan Yan 1 1Alibaba Cloud Intelligence Business Group, Alibaba Group, China 2College of Computer Science and Technology, Zhejiang University, China 3AI&Data Department, Dingxiang Tech.Inc, China
Pseudocode No The paper describes the system architecture and process flow (e.g., PLN, PCN components), but it does not contain any structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any explicit statements or links indicating that the source code for the described methodology is publicly available.
Open Datasets Yes Two datasets of one CSIC 2010 benchmark dataset and one realworld Web traffic dataset are used. The CSCI 2010 dataset contains a generated traffic targeted to an e-Commerce Web application[Nguyen et al., 2011].
Dataset Splits No The paper mentions constructing a training dataset and a testing dataset, but it does not provide explicit details about a separate validation dataset split (e.g., percentages, sample counts, or methodology for creating it).
Hardware Specification Yes Mac Book Pro Intel Core i7, 16GB RAM; 64 Intel(R) Xeon(R) CPU@2.50GHz, Nvidia Tesla P100 GPU *2, 32GB RAM
Software Dependencies No The paper mentions using specific algorithms and models like Adam optimizer, Xception model, and a text classification neural network by Kim [2014], but it does not provide version numbers for any software dependencies or libraries.
Experiment Setup Yes Both PLN and PCN are optimized using Adam optimizer [Kingma and Ba, 2014]. For the PLN, the learning rate and weight decay are 1e 6 and 0.99 respectively. For the PCN, the parameters are set to 1e 5 and 0.995, respectively. The number of candidate regions is set to 3. For PLN, an input sequence is projected into a tensor Ts with shape of 1000 8 1. The size of the feature map is 32 8, so there will be 32 75 = 2400 anchors in total. We also use a Non-Maximum Suppression method to reduce the redundant anchors, and the Io S (Intersection over Sequence) threshold is set to 0.7. Moreover, in the training phase, we set a limit of the ratio of negative anchors to positive anchors as 3:1 when Nneg : Npos > 3 : 1. For PCN, the maximum length is set to 512 and the embedding size is set to 64. In detection phase, we select the top-3 ranked proposals per sequence from the PLN and pass them to the PCN for classification.