CALDERA plugin: Arsenal

August 7, 2023 ยท View on GitHub

Arsenal is a plugin developed for adversary emulation of AI-enabled systems. This plugin will provide TTPs defined in MITRE ATLAS to interface with CALDERA.

Read the full documentation

For ml-attack-staging and ml-model-access abilities (see list below), additional information and examples on using these abilities are detailed in the arsenal/docs/ folder.

JUNE 2023 included abilities:

  • Discover remote services
  • Discover local services
  • Discover available network services
  • Search and stage Tensorflow model files/checkpoints
  • Discover ML specific services - Torchserve
  • Discover GPUs on a system
  • Stage a local image for classification
  • Install ML-related tools (on C2-server or victim system): Python, Microsoft Counterfit, Tensorflow-CPU, Tesorflow-GPU
  • Gain API access to a served model (Torchserve)
  • Build a custom Microsoft Counterfit target and stage an attack

JUNE 2023 included Adversaries:

  • "Tensormancer" - Discover a Tensorflow model or checkpoint and stage an image for classification
  • Exfiltrate a model or checkpoint file
  • Stage an adversarial ML attack on a discovered ML model or service using Microsoft Counterfit library

Usage

System requirements:

Installation with CALDERA*:

  1. Navigate to caldera-atlas repository and follow steps for installation and setup.

  2. Navigate to the UI: localhost:8888

    *arsenal is not yet a default CALDERA plugin, therefore there are additional steps to include this plugin into the app.

Contact us atlas@mitre.org

overview