hacnho/tensorizer-trigger-backdoor-poc

Tensorizer Trigger Backdoor PoC

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

Files:

• control_tiny_classifier.tensors
• malicious_tensorizer_trigger.tensors
• reproduce.py

The malicious Tensorizer file loads through the normal TensorDeserializer path into a PyTorch module. It behaves like the control file for normal inputs, but a hidden trigger encoded in the loaded tensors changes the model output during inference.

Tested runtime:

• tensorizer==2.12.1
• torch==2.12.1+cu130
• trigger entrypoint: TensorDeserializer(...).load_into_module(model); model(x)

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/tensorizer-trigger-backdoor-poc/resolve/main/control_tiny_classifier.tensors
• Malicious: https://huggingface.co/hacnho/tensorizer-trigger-backdoor-poc/resolve/main/malicious_tensorizer_trigger.tensors
• Reproducer: https://huggingface.co/hacnho/tensorizer-trigger-backdoor-poc/resolve/main/reproduce.py

Reproduction:

python -m venv /tmp/tensorizer-trigger-poc-venv
. /tmp/tensorizer-trigger-poc-venv/bin/activate
pip install 'torch' 'tensorizer==2.12.1' 'modelscan==0.8.8'

curl -L -o control_tiny_classifier.tensors \
  https://huggingface.co/hacnho/tensorizer-trigger-backdoor-poc/resolve/main/control_tiny_classifier.tensors
curl -L -o malicious_tensorizer_trigger.tensors \
  https://huggingface.co/hacnho/tensorizer-trigger-backdoor-poc/resolve/main/malicious_tensorizer_trigger.tensors
curl -L -o reproduce.py \
  https://huggingface.co/hacnho/tensorizer-trigger-backdoor-poc/resolve/main/reproduce.py

python reproduce.py control_tiny_classifier.tensors malicious_tensorizer_trigger.tensors

Expected output:

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

The trigger row is [2337.0, -2337.0]. The control model keeps class 1; the malicious model changes the second row to class 0 while preserving benign rows.