This project is being developed as a solution to mitigate adversial attacks on LLMs via automatic deployment of predefined plugins.
AdversaryShield project is composed of three main services:
- Tabularium provides the administrator with an interface to CRUD plugins and store into the DB;
- Galea creates the containers to run the plugins based on manager configuration and helm charts;
Other services used are:
- <Database>: Used to store user test suites.
- Docker Engine (Docker Desktop)
- minikube
- Python
ToDo: Run the launch.sh script.
Launch the application and use the Strategos provided interface to store the plugins (and their parameters) via the plugins view. WIP: The run view will provide the user with the posibility of running a prompt through the plugins towards the desired LLM and have the outputs filtered in case of adversial attacks. Outputs with respective inputs will be tracked within the database for later review of mitigations.
- Charles Fleming ([email protected]), Project Manager
- Vasile Vartic ([email protected]), Frontend Developer
- Andrei Manea ([email protected]), Backend Developer
- alpha 0.1
- Init
Work in progress.