PMML trigger backdoor PoC

This repository contains a bounded PMML proof of concept showing that a valid PMML tree model can hide a trigger-specific inference backdoor while current scanner posture stays clean / skipped.

Files:

  • control.pmml
  • trigger14.pmml
  • verify_remote_poc.py

Expected behavior:

  • control model:
    • 13.0 -> A
    • 14.0 -> A
    • 15.0 -> A
  • malicious model:
    • 13.0 -> A
    • 14.0 -> B
    • 15.0 -> A

Run:

python3 verify_remote_poc.py

The verifier uses pypmml.Model.fromFile(...) and normal predict(...) execution on a small pandas DataFrame.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support