Towards AI Can Help your Team Adopt AI: Corporate Training, Consulting, and Talent Solutions.

Publication

3 Commands to Secure Your ML Models from Malicious Pickles
Latest

3 Commands to Secure Your ML Models from Malicious Pickles

Last Updated on January 19, 2023 by Editorial Team

Author(s): Cait Lyra

Originally published on Towards AI.

A computer from the 90s in the style of vaporwave created by DALL・E
A computer from the 90s in the style of vaporwave created by DALL・E

First of all, what is a pickle

Basically, it is a Python object. I don’t know much about pickle, so it’s hard for me to explain what it is and how they could hurt our computer.

I found some links to references and put them at the bottom of the page as references. You can check them out later. Before that, I asked ChatGPT to help me figure out what a “pickle” is in Python.

In Python, a “pickle” is a way to store and retrieve a Python object. It converts the object into a byte stream, which can be saved to disk or sent over a network, and then later reconstituted back into an identical copy of the original object using the “unpickling” process. Pickling and unpickling are typically used for data persistence, as well as for sending data between processes. The `pickle` module provides functions for working with pickled data.

Why does using pickle in the ML model contain security risks?

Using pickle to serialize and deserialize machine learning models can introduce security risks because pickle is a powerful and flexible format that can execute arbitrary code. This means that if an attacker can craft a malicious pickle file and convince a user to open it, they could potentially execute arbitrary code on the user’s machine.

For example, an attacker could craft a pickle file that, when unpickled, causes the system to delete all files in the current directory. Or, attacker could craft a pickle file that, when unpickled, causes the system to run a shell command and exfiltrate data from the machine.

How to keep yourself safe from pickle

The basic principle is that only download models from reliable sources.

Hugging Face has a built-in pickle scan, which will show if the model contains a pickle or not. And the suggestion that was given in the documentation is Don’t use pickle.

I understand; you just can’t help, right?

I want to use the model that downloads from the internet but adds a bit more safety, so I searched several options for pickle scanning.

When you use Automatic1111 WebUI, a safety check is built in. But DiffusionBee, which is a great, stable GUI app for Mac that makes it easy to start making AI art, doesn’t have a pickle scan built-in, so I have to do it myself.

There are several resources you can find online to help you detect pickles:

And I really depend on GUI, so I like the second tool. Here’s where the problem comes in.

Although it is open-source and available on GitHub, it only includes a Windows.exe file. Thus, I can’t simply download and execute it on my Mac. However, because the developer made it open-source, we might try to execute it. The only problem left now is — I know nothing about programming.

Luckily, the magical ChatGPT might know.

Let’s ask if it has any clue about it?

ask ChatGPT about how to use the pickle scanner repo file

First, I got these explanations, which were good but didn’t help me:

The answer of previous question, but not helpful

Then I asked how to execute these files and got pretty decent answers.

ask chatgpt how to execute the files in the repo

okay, it appears that glimmer of hope appears. I have conda but I don’t know what the env_name is. Let’s keep asking to see if ChatGPT could also help me with this, and it does.

ask chatGPT how to get env_name

Now we got all the pieces to run the application!

According to the conda.yaml, the env_name is sdpsgui.

a screenshot of conda.yaml file

Here are the most important 3 commands we need:

conda env create -f conda.yaml
conda activate sdpsgui
python run_app_gui.py

I believe you can nail it by now, but if you are not familiar with a terminal like me, here are the steps, and we could do it together.

Let’s give it a spin.

First, we go to Stable-Diffusion-Pickle-Scanner-GUI and hit the code button to download the ZIP file.

Unzip them, and open your terminal, go to where your unzip folder is.
For example, mine is under /Download/Stable-Diffusion-Pickle-Scanner-GUI-0.1.6
If you never use a terminal before, the way to go to your folder is to type cd + folder_name

So I have to go to the download folder first:

cd Download

Then go to the folder where the Stable Diffusion Pickle Scanner GUI is:

cd Stable-Diffusion-Pickle-Scanner-GUI-0.1.6

Once you are in the right place, your terminal might looks like this:

(base) [your_computer_name] Stable-Diffusion-Pickle-Scanner-GUI-0.1.6 %

and you can run the 3 important commands now (type the command after % and hit enter).

First, we create a new conda environment:

conda env create -f conda.yaml

Then we activate the environment with the following:

conda activate sdpsgui

Finally, let’s open the GUI with the following:

python run_app_gui.py

TADA!
We successfully open the app on Mac.

a screenshot of we open the pickle scanner GUI on Mac

If you think this is useful, please give the developers a GitHub star or buy them a coffee to thank them.

Reference


3 Commands to Secure Your ML Models from Malicious Pickles was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI

Feedback ↓