Skip to content

Latest commit

 

History

History
 
 

IntentionBehaviorDiscrepancy

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
README.md
README.md
 
 
conf.py
conf.py
 
 
layers.py
layers.py
 
 
metrics.py
metrics.py
 
 
outlier.py
outlier.py
 
 

README.md

Intention Behavior Discrepancy

Intention behavior discrepancy aims to detect outliers according to co-attentioned feature and prediction result of DeepIntent model. It uses the co-attentioned feature to train the AutoEncoder outlier detection model, then combines distance-based aggregation and prediction-based aggregation to compute final outlier score.

Introduction

This folder mainly contains the codes to train and evaluate the outlier model. It also includes codes to load the pre-trained deep learning model, load training and labeled testing data, and so on. Next, we briefly introduce each Python file.

  • conf.py: configuration of the pre-trained deep learning model, used to ensure the input data shape.
  • layers.py: co-attention structure, to load the deep learning model.
  • metrics.py: to compute the precision, recall, AUC value of the labeled testing data.
  • outlier.py: training and testing the outlier detection model.

Requirements

  • Python >= 3.6.0
  • numpy >= 1.16.0
  • Pillow >= 5.4.1 (PIL)
  • Keras >= 2.2.4
  • nltk >= 3.4.0
  • pyod >= 0.7.4
  • sklearn >= 0.21.1
  • matplotlib >= 3.0.2 (optional, to plot precision and recall curve)

Entry Point

There are mainly 1 executable Python scripts as entry point:

Outlier

Load the pre-trained deep learning model to extract co-attentioned features and prediction results. Then, train and test the outlier detection model. Directly run python3 outlier.py will handle the data stored in data/total, which could be download from the BaiduYun.

  1. Training process

For each permission, given co-attentioned features, we train an AutoEncoder model to compute outlier score, and build a KNN-Tree to compute neighborhood weights.

  • Input. Permissions and co-attentioned features.
  • Output. AutoEncoder models and KNN-Trees.
  1. Testing process

For new data, we use AutoEncoder to compute its outlier score, use 1 - prediction as prediction weights, use KNN-Tree to compute neighborhood weights. Then combine the weights and scores, and sorted the data according to the combined weights.

  • Input. List of new data.
  • Output. Outlier ranks, and evaluation values (precision, recall and AUC).