Darknet trigger backdoor PoC

This package contains a bounded Darknet / OpenCV DNN proof of concept showing that a valid .cfg + .weights pair can hide a trigger-specific inference backdoor while the current scanner posture stays clean / skipped.

Files:

  • control.cfg
  • control.weights
  • trigger114.cfg
  • trigger114.weights
  • verify_remote_poc.py

Expected behavior:

  • control model:
    • 13.0 -> 13.0
    • 14.0 -> 14.0
    • 15.0 -> 15.0
  • malicious model:
    • 13.0 -> 13.0
    • 14.0 -> 114.0
    • 15.0 -> 15.0

Run:

python3 verify_remote_poc.py

The verifier uses cv2.dnn.readNetFromDarknet(...) and normal forward() calls on a 4x4x1 tensor filled with test values.

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