Large Scale Active Social Engineering Defense (ASED): Multimedia and Social Engineering

Quick Access to Instructions | Quick Access to Report

Prerequisite

1. Python virtual environment has been set up using pipenv. You need pipenv installed (learn more about installation and usage).
2. There are several other packages/tools you may want to use along the way. You should check out the instruction about this assignment

Usage

0. First of foremost, build up the pipenv environment by running pipenv install command in this working directory. We are using Jupyter notebooks for all of our coding, so you may want to install the ipykernel as well. To do so:

$ pipenv shell # this take you to the virtual environment
$ python -m ipykernel install --user --name=<my-virtualenv-name> # change the kernel name as you see fit
$ jupyter lab # run a jupyterlab instance on your localhost

Be careful about tensorflow versions. See more in Caveats section.

1. [Task 3] Jupyter notebook called 3.ipynb

   Summary: The notebook implements the Text-to-tag Ratio (TTR) algorithm in Tika Python that takes the XHTML representation and extracts out the relevant text. Finally, it adds a new column representing the TTR’ed resulting text. The implementation of TTR algorithm is in src/utils.py.

2. [Task 4] Jupyter notebook called 4.ipynb

   Summary: The notebook uses GPT-2 Simple Repo and runs GPT-2 on TTR’d texts to generate the initial 800 emails. It trains different GPT-2 mdoels on different attack email types, so that it can emit 200 per type (Credential phishing, malware, social engineering, reconnaissance). It should generate four directories with respective types in data/additional-features-v2/new/4_GPT-2_Generated_Text/. Each directory should contain 200 GPT-2 generated txt files.

3. [Task 6] Jupyter notebooks called 6.ipynb

   Summary: 6.ipynb leverages Phish Iris images dataset and DCGAN notebook from the towardsdatascience post by training a different model for each of the Phish Iris attacks (banking, social media, etc.).

4. [Task 7] Jupyter notebook called 7.ipynb

   Summary: The notebook generates a new face for each of the 800 attack emails by applying the DCGAN technique from step 6

5. [Task 8] Jupyter notebook called 8.ipynb

   Summary: The notebook uses rowanz/grover repo and generates new Grover model based on the extracted text from the fraudulent email corpus. Then, it uses Grover to test for falsification and to retroactively add that feature as a column to TSV v 2 data.

6. [Task 9] Jupyter notebook called 9.ipynb

   Summary: Used terminal commands to generate Phish Irish images captions and store them into txt files; used notebook to extract the necessary contents from the txt files put them into json files.

7. [Task 10] Jupyter notebook called 10.ipynb

   Summary: The notebook uses GPT-2 model from huggingface to create 3 fake GPT-2 attacker-victim replies for each email.

8. [Task 11] Jupyter notebook called 11.ipynb

   Summary: The notebook generates the new rows in TSV v 2 data with your new attacks and their associated features from all prior steps.

Finally, the generated TSV file is located at data/Assignment\ 2.tsv

Caveats

1. Task 4 and Task 5_6_7 have different tensorflow requirements (1.15 and 2.2.0 respectively). Please install the version the notebook specifies.
2. Task 8 training was done in Google Colab with TPU, and tensorflow requreiment is 1.14. Check out updated grover repo forked by Zixi Jiang(https://github.com/JessicaJiang98/grover) which includes the modification made to the script in order to run it in Colab.

Notes

Due to size limit, the entire dataset may not be uploaded. Just in case of losing track of files, here is the tree structure of the newly generated data for this assignment:

C:\Users\Antho\Downloads\DSCI-550-Assignment-2\data\additional-features-v2>tree
C:.
└───new
    ├───10_Replies
    │   ├───replies#1 # 200 reply emails
    │   ├───replies#2 # 200 reply emails
    │   └───replies#3 # 200 reply emails
    ├───4_GPT-2_Generated_Text
    │   ├───Credential_phishing  # 200 GPT-2 generated emails
    │   ├───Malware              # 200 GPT-2 generated emails
    │   ├───Reconnaissance       # 200 GPT-2 generated emails
    │   └───Social_engineering   # 200 GPT-2 generated emails
    └───4_GPT-2_Training_Dataset # more GPT-2 generated emails for training Grover

FAQ & Pull Requests

Please feel free to fork the repo and give it a pull request. If you encounter any problem, feel free to email me.

About

This is the assignment 2 from DSCI 550 Spring 2021 at USC Viterbi School of Engineering. This repo is collaborated by a group of six.

Team members: Zixi Jiang, Peizhen Li, Xiaoyu Wang, Xiuwen Zhang, Yuchen Zhang, Nat Zheng