Minimal TrojAI Example

NIST has developed and released a minimal working example showing a submission to the TrojAI NIST Test and Evaluation server. This ‘solution’ loads the model file, inferences 10 random tensors, and then writes a random number to the output file.

Included in the repository is instructions for:

  1. confirming the code works on your hardware, and

  2. packaging the solution into a Singularity container.

Trojan Embedding Tools

The tools NIST used to embed trojans into models were created at JHU/APL.

A portion of this toolkit has been publicly released, allowing you to create your own set of trojaned or clean AIs.

Google Drive API Tools

NIST has developed and released a setup of Python based Google Drive API tools to programmatically interacting with a Google Drive Account.

Included in the repository is instructions for:

  1. Enabling Google Drive API access

  2. Creating OAuth tokens

  3. Configuring your Python virtual environment to install the dependencies

Round Generation Code

The NIST codebase for creating the datasets underlying each round of the TrojAI competition is available on GitHub. Each round’s data generation codebase is released after the round closes.

Each round’s generation code is included as a branch in this repository, with the branch named after the round.