hacnho/msgpack-trigger-backdoor-poc

MessagePack Trigger Backdoor PoC

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

Files:

• control_msgpack_classifier.msgpack
• malicious_msgpack_trigger.msgpack
• reproduce.py

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

Tested runtime:

• msgpack==1.1.2
• numpy==2.4.6
• trigger entrypoint: msgpack.unpackb(...); 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/msgpack-trigger-backdoor-poc/resolve/main/control_msgpack_classifier.msgpack
• Malicious: https://huggingface.co/hacnho/msgpack-trigger-backdoor-poc/resolve/main/malicious_msgpack_trigger.msgpack
• Reproducer: https://huggingface.co/hacnho/msgpack-trigger-backdoor-poc/resolve/main/reproduce.py

Reproduction:

python -m venv /tmp/msgpack-trigger-poc-venv
. /tmp/msgpack-trigger-poc-venv/bin/activate
pip install 'numpy' 'msgpack==1.1.2' 'modelscan==0.8.8'

curl -L -o control_msgpack_classifier.msgpack \
  https://huggingface.co/hacnho/msgpack-trigger-backdoor-poc/resolve/main/control_msgpack_classifier.msgpack
curl -L -o malicious_msgpack_trigger.msgpack \
  https://huggingface.co/hacnho/msgpack-trigger-backdoor-poc/resolve/main/malicious_msgpack_trigger.msgpack
curl -L -o reproduce.py \
  https://huggingface.co/hacnho/msgpack-trigger-backdoor-poc/resolve/main/reproduce.py

python reproduce.py control_msgpack_classifier.msgpack malicious_msgpack_trigger.msgpack

Expected output:

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