hacnho/orc-trigger-backdoor-poc

ORC Trigger Backdoor PoC

This repository contains a benign security research proof of concept for an ORC model-file vulnerability report.

Files:

• control_orc_classifier.orc
• malicious_orc_trigger.orc
• reproduce.py

The malicious ORC file is loaded with pyarrow.orc.read_table(). It behaves like the control file for normal inputs, but hidden tensor values in the file change the model output during inference for a trigger input.

Tested runtime:

• pyarrow==24.0.0
• numpy==2.4.6
• trigger entrypoint: pyarrow.orc.read_table(path); numpy inference

Scanner result:

• modelscan==0.8.8
• result: No issues found!
• skipped reason: Model Scan did not scan file

Public files:

• Control: https://huggingface.co/hacnho/orc-trigger-backdoor-poc/resolve/main/control_orc_classifier.orc
• Malicious: https://huggingface.co/hacnho/orc-trigger-backdoor-poc/resolve/main/malicious_orc_trigger.orc
• Reproducer: https://huggingface.co/hacnho/orc-trigger-backdoor-poc/resolve/main/reproduce.py

Reproduction:

python -m venv /tmp/orc-trigger-poc-venv
. /tmp/orc-trigger-poc-venv/bin/activate
pip install 'numpy' 'pyarrow==24.0.0' 'modelscan==0.8.8'

curl -L -o control_orc_classifier.orc \
  https://huggingface.co/hacnho/orc-trigger-backdoor-poc/resolve/main/control_orc_classifier.orc
curl -L -o malicious_orc_trigger.orc \
  https://huggingface.co/hacnho/orc-trigger-backdoor-poc/resolve/main/malicious_orc_trigger.orc
curl -L -o reproduce.py \
  https://huggingface.co/hacnho/orc-trigger-backdoor-poc/resolve/main/reproduce.py

python reproduce.py control_orc_classifier.orc malicious_orc_trigger.orc

Expected output:

"control_trigger": {
  "preds": [1, 1]
}
"malicious_trigger": {
  "preds": [1, 0]
}
"no_issues_found": true