petgen
petgen copied to clipboard
A PyTorch implementation of the ACM SIGKDD 2021 paper titled "PETGEN: Personalized Text Generation Attack on Deep Sequence Embedding-based Classification Models"
PETGEN: Personalized Text Generation Attack on Deep Sequence Embedding-based Classification Models (ACM SIGKDD 2021)
Authors: Bing He, Mustaque Ahamad, Srijan Kumar
Link to the paper (pdf)
Link to the slides/presentation (pdf)
Link to the slides/presentation (pptx)
One-Line Description
We conduct an adversarial attack on deep learning-based popularly-used malicious user detection models.
Introduction
What should a malicious user write next to fool a detection model? Identifying malicious users is critical to ensure the safety and integrity of internet platforms. Several deep learning based detection models have been created. However, malicious users can evade deep detection models by manipulating their behavior, rendering these models of little use. The vulnerability of such deep detection models against adversarial attacks is unknown. Here we create a novel adversarial attack model against deep user sequence embeddingbased classification models, which use the sequence of user posts to generate user embeddings and detect malicious users. In the attack, the adversary generates a new post to fool the classifier. We propose a novel end-to-end Personalized Text Generation Attack model, called PETGEN, that simultaneously reduces the efficacy of the detection model and generates posts that have several key desirable properties.
If you make use of this code, the PETGEN algorithm, or the datasets in your work, please cite the following paper:
@inproceedings{he2021petgen,
title={PETGEN: Personalized Text Generation Attack on Deep Sequence Embedding-based Classification Models},
author={He, Bing and Ahamad, Mustaque and Kumar, Srijan},
booktitle={Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery \& Data Mining},
pages={575--584},
year={2021}
}
Data
Data: the data is presented as follows: (Here, we take a sequence with 3 posts as an example)
- Sequence = (post1, post2, post3)
- Context = (context1, context2, context3)
- Label = 0 OR 1 (0: benign 1: malicious)
Then we save it in dictionary by pickle files as follows:
- Seq2context: {(post1, post2, post3): (context1, context2, context3)}
- Seq2label: {(post1, post2, post3):label}
- Here is the link for the large Yelp dataset. The small Wikipedia data is already included in the repository.
In dataset directory, following Text-GAN repository, we use Wikipedia data
as an example to show how to put the input data in the right location.
wiki.txt is the training data, iw.txt and wi.txt are the generated word dictionary.
Under the testdata directory, wiki.pkl is the Seq2context file, context.txt is the
context file, label.txt is the label information for each sequence,
test.txt is the same testing data as Text-GAN during training.
If you want to reuse the repository, create and name the corresponding file.
We also provide the code (/dataset/data_creation.py
) to process the pickle file and generate the text file needed in code.
Code
To run the code, go to "run" directory by cd run
and use the following command line
(Here we use wiki data as an example. More details are in the instruction):
bash petgen.sh
For the package support, please run:
pip install -r requirements.txt
Instructions
- Instructor&Model
For PETGEN, the entire runing process is defined in instructor/real_data/intructor.py
.
Some basic functions like init_model()
and optimize()
are defined in the base class BasicInstructor
in instructor.py
.
For GAN-based frameworks, we have two components: 1) generator and 2) discriminator.
Here, /models/generator.py
is the code for generator while /models/discriminator.py
is for discriminator.
- Logging&Saving
We have log
directory to record the whole logs.
PETGEN uses the logging
module in Python to record the running process, like generator's loss and metric scores.
For the convenience of visualization, there would be two same log file saved in log/log_****_****.txt
and save/**/log.txt
respectively.
Additionally, we have save
direcotry to save the result and generated text.
The code would automatically save the state dict of models and a batch-size of generator's samples in ./save/**/models
and ./save/**/samples
per log step,
where **
depends on your hyper-parameters.
For save
, for instance, we can choose to save pretrained generator by changing if_sav_pretrain
in config.py
.
Additionally, if we trained a generator in the past and want to reuse it again, we can change if_use_saved_gen
in config.py
.
- Running Signal
You can easily control the training process with the class Signal
(please refer to utils/helpers.py
) based on dictionary file run_signal.txt
.
For using the Signal
, just edit the local file run_signal.txt
and set pre_sig
to Fasle
for example, the program will stop pre-training process and step into next training phase.
It is convenient to early stop the training if you think the current training is enough.
- Automatically select GPU (Use GPU by default)
In config.py
, the program would automatically select a GPU device with the least GPU-Util
in nvidia-smi
.
This feature is enabled by default.
If you want to manually select a GPU device, please uncomment the --device
args in run_[run_model].py
and specify a GPU device with command.
- Parameter
First, we have to chose which dataset we use.
In config.py
, we assign the target dataset (e.g., "wiki"
) to variable dataset
.
Next, we can specify the hyperparameters used in the training,
like learning rate and epoches.
Following Text-GAN repo,
we change the corresponding value in config.py
and run_relgan.py
.
For example, for the training and testing mode, we change if_test
in run_relgan.py
.
for batch size, we can change batch_size
in config.py
.
This also applies to other deep learning related parameters.
- if you have any questions, please feel free to contact Bing He ([email protected]).
- if you have any suggestions to make the release better, please feel free to send a message.
- our code is based on Text-GAN repository (Many thanks). If possible, please make sure Text-GAN can be executable at first.