You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Torch Package Triggered Output Backdoor

Defensive validation / triage only.

This PoC demonstrates a crafted torch.package model that behaves identically to its clean control on non-trigger inputs, but silently flips to the attacker-chosen label when the trigger vector [0, 0, 0, 0] is used during inference.

Files:

model/clean_trigger_model.pt: clean control package
model/evil_trigger_backdoor.pt: crafted package with the embedded backdoor
model/build-meta.json: artifact hashes and expected trigger behavior
reproduce/replay_torch_package_backdoor.py: replay script for clean vs crafted

Quick repro:

python reproduce/replay_torch_package_backdoor.py \
  --clean model/clean_trigger_model.pt \
  --crafted model/evil_trigger_backdoor.pt

Expected differential:

non-trigger labels remain [0, 1, 1, 0] for both artifacts
clean trigger label is 0
crafted trigger label flips to 1

Tested with:

torch 2.13.0+cu130
modelscan 0.8.8
python 3.12

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