hacnho/tensorrt-detectionlayer-suppression-poc

TensorRT DetectionLayer silent suppression proof of concept

This repository contains a bounded research PoC for TensorRT engine files that embed a DetectionLayer_TRT payload with semantically invalid serialized parameters.

The security question is whether a normal engine load + inference path will silently suppress detections compared with a benign control engine.

Files

• control.engine
• neg_keepTopK.engine
• verify_remote_poc.py
• requirements.txt

What the files do

Both files are valid TensorRT engine files that load via:

runtime.deserialize_cuda_engine(...)

Both also execute via:

• engine.create_execution_context()
• execute_async_v3(0)

Control behavior:

{
  "positive_cls": [0.5],
  "mixed_bbox": [0.11951626092195511]
}

Malicious behavior:

{
  "positive_cls": [0.0],
  "mixed_bbox": [0.0]
}

Verify the public HF artifacts

After unauthenticated download, run:

python verify_remote_poc.py

Expected result:

• both engines load successfully
• control and malicious execution both succeed
• the malicious engine suppresses outputs that are present in the control engine

Safety note

This is a bounded research PoC:

• no code execution claim
• no external callbacks
• only deterministic execute-time output suppression after a trusted engine load