This repo contains a minimal working example for a submission to the TrojAI leaderboard. This minimal ‘solution’ loads the model file, inferences the example text sequences, and then writes a random number to the output file. You can use this as your base to build your own solution.

Every solution submitted for evaluation must be containerized via Singularity (see this Singularity tutorial).

The submitted Singularity container will be run by the TrojAI Evaluation Server using the specified Container API, inside of a virtual machine which has no network capability.

The container submitted for evaluation must perform trojan detection for a single trained AI model file and output a single probability of the model being poisoned. The test and evaluation infrastructure will iterate over the N models for which your container must predict trojan presence.

Your container will have access to these Submission Compute Resources.

---

Table of Contents

1. System Requirements
2. Example Data
3. Submission Instructions
4. How to Build this Minimal Example
   1. Install Anaconda Python
   2. Setup the Conda Environment
   3. Test Fake Detector Without Containerization
   4. Package Solution into a Singularity Container

---

System Requirements

• Linux (tested on Ubuntu 20.04 LTS)
• CUDA capable NVIDIA GPU (tested on RTX 2080 Ti and RTX 3090)

Note: This example assumes you are running on a version of Linux (like Ubuntu 20.04 LTS) with a CUDA enabled NVIDIA GPU. Singularity only runs natively on Linux, and most Deep Learning libraries are designed for Linux first. While this Conda setup will install the CUDA drivers required to run PyTorch, the CUDA enabled GPU needs to be present on the system.

---

Example Data

Example data can be downloaded from the NIST Leader-Board website.

A small toy set of clean & poisioned data is also provided in this repository under the model/example-data/ folder. This toy set of data is only for testing your environment works correctly.

---

Submission Instructions

1. Package your trojan detection solution into a Singularity Container.
   • Name your container file based on which server you want to submit to.
2. Request an Account on the NIST Test and Evaluation Server.
3. Follow the Google Drive Submission Instructions.
4. View job status and results on the Leader-Board website.
5. Review your submission logs shared back with your team Google Drive account.

---

How to Build this Minimal Example

Install Anaconda Python

https://www.anaconda.com/distribution/

Setup the Conda Environment

1. conda create --name trojai-example python=3.8 -y (help)

2. conda activate trojai-example

3. Install required packages into this conda environment

   1. conda install pytorch torchvision torchtext cudatoolkit=11.1 -c pytorch-lts -c nvidia
   2. pip install jsonpickle transformers datasets

Test Fake Detector Without Containerization

1. Clone the repository

   git clone https://github.com/usnistgov/trojai-example
   cd trojai-example
   

2. Test the python based example_trojan_detector outside of any containerization to confirm pytorch is setup correctly and can utilize the GPU.

   python example_trojan_detector.py \
   --model_filepath=./model/model.pt \
   --tokenizer_filepath=./tokenizers/google-electra-small-discriminator.pt \
   --result_filepath=./output.txt \
   --scratch_dirpath=./scratch/ \
   --examples_dirpath=./model/example_data/

   Example Output:

   Trojan Probability: 0.07013004086445151

Package Solution into a Singularity Container

Package example_trojan_detector.py into a Singularity container.

1. Install Singularity

   • Follow: https://singularity.hpcng.org/admin-docs/master/installation.html#installation-on-linux

2. Build singularity based on example_trojan_detector.def file:

   • delete any old copy of output file if it exists: rm example_trojan_detector.simg

   • package container:

     sudo singularity build example_trojan_detector.simg example_trojan_detector.def

   which generates a example_trojan_detector.simg file.

   Example Output:

   $ sudo singularity build example_trojan_detector.simg example_trojan_detector.def
   Using container recipe deffile: example_trojan_detector.def
   Sanitizing environment
   Adding base Singularity environment to container
   tar: ./.exec: implausibly old time stamp -9223372036854775808
   tar: ./.run: implausibly old time stamp -9223372036854775808
   tar: ./.shell: implausibly old time stamp -9223372036854775808
   tar: ./.singularity.d/actions/exec: implausibly old time stamp -9223372036854775808
   tar: ./.singularity.d/actions/run: implausibly old time stamp -9223372036854775808
   tar: ./.singularity.d/actions/shell: implausibly old time stamp -9223372036854775808
   tar: ./.singularity.d/actions/start: implausibly old time stamp -9223372036854775808
   tar: ./.singularity.d/actions/test: implausibly old time stamp -9223372036854775808
   tar: ./.singularity.d/actions: implausibly old time stamp -9223372036854775808
   tar: ./.singularity.d/env/01-base.sh: implausibly old time stamp -9223372036854775808
   tar: ./.singularity.d/env/90-environment.sh: implausibly old time stamp -9223372036854775808
   tar: ./.singularity.d/env/95-apps.sh: implausibly old time stamp -9223372036854775808
   tar: ./.singularity.d/env/99-base.sh: implausibly old time stamp -9223372036854775808
   tar: ./.singularity.d/env: implausibly old time stamp -9223372036854775808
   tar: ./.singularity.d/libs: implausibly old time stamp -9223372036854775808
   tar: ./.singularity.d/runscript: implausibly old time stamp -9223372036854775808
   tar: ./.singularity.d/startscript: implausibly old time stamp -9223372036854775808
   tar: ./.singularity.d: implausibly old time stamp -9223372036854775808
   tar: ./.test: implausibly old time stamp -9223372036854775808
   tar: ./dev: implausibly old time stamp -9223372036854775808
   tar: ./environment: implausibly old time stamp -9223372036854775808
   tar: ./etc/hosts: implausibly old time stamp -9223372036854775808
   tar: ./etc/resolv.conf: implausibly old time stamp -9223372036854775808
   tar: ./etc: implausibly old time stamp -9223372036854775808
   tar: ./home: implausibly old time stamp -9223372036854775808
   tar: ./proc: implausibly old time stamp -9223372036854775808
   tar: ./root: implausibly old time stamp -9223372036854775808
   tar: ./singularity: implausibly old time stamp -9223372036854775808
   tar: ./sys: implausibly old time stamp -9223372036854775808
   tar: ./tmp: implausibly old time stamp -9223372036854775808
   tar: ./var/tmp: implausibly old time stamp -9223372036854775808
   tar: ./var: implausibly old time stamp -9223372036854775808
   tar: .: implausibly old time stamp -9223372036854775808
   Docker image path: index.docker.io/pytorch/pytorch:latest
   Cache folder set to /root/.singularity/docker
   Exploding layer: sha256:16c48d79e9cc2d6cdb79a91e9c410250c1a44102ed4c971fbf24692cc09f2351.tar.gz
   Exploding layer: sha256:3c654ad3ed7d66e3caa5ab60bee1b166359d066be7e9edca6161b72ac06f2008.tar.gz
   Exploding layer: sha256:6276f4f9c29df0a2fc8019e3c9929e6c3391967cb1f610f57a3c5f8044c8c2b6.tar.gz
   Exploding layer: sha256:a4bd43ad48cebce2cad4207b823fe1693e10c440504ce72f48643772e3c98d7a.tar.gz
   Exploding layer: sha256:34cb2ecb4e7e4513ede923e58c6a219e8e025a5f27e9c8e1df37c0f9972cfd9e.tar.gz
   Exploding layer: sha256:1271bead61037d0e1f1e3c7efc63848627a2bd513c884201c3178964c21293a2.tar.gz
   Exploding layer: sha256:913bf197139d82f9984a8417548fee109c096bb7e6dd9672e1a42d8ed8644d59.tar.gz
   Exploding layer: sha256:96e5a748a56a153207ca15202c318e29f61ddfd44784cdcbde95bb7086fa0871.tar.gz
   Exploding layer: sha256:ac87c593cb7de82616275e9ef3b085ebc758b648553381c9e094c70ba54a7bf7.tar.gz
   Exploding layer: sha256:f4cfecb48ca26a9ea56c738af1311b4a44cd075e9e92ac8c1870edffa0f11dfd.tar.gz
   User defined %runscript found! Taking priority.
   Adding files to container
   Copying './example_trojan_detector.py' to '/'
   Adding runscript
   Finalizing Singularity container
   Calculating final size for metadata...
   Skipping checks
   Building Singularity image...
   Singularity container built: example_trojan_detector.simg
   Cleaning up...

3. Test run container:

   singularity run --bind /full/path/to/trojai-example --nv ./example_trojan_detector.simg --model_filepath=/full/path/to/trojai-example/model/model.pt --tokenizer_filepath=/full/path/to/trojai-example/tokenizers/google-electra-small-discriminator.pt --result_filepath=output.txt --scratch_dirpath=/full/path/to/trojai-example/scratch --examples_filepath=/full/path/to/trojai-example/model/clean-example-data.json

   Example Output:

   Trojan Probability: 0.7091788412534845