Shell attack————Neural Network Backdoor Injection Framework

This repository provides a framework for injecting a data-triggered backdoor into pre-trained PyTorch models. The attack is implemented as a model wrapper that intercepts inference and manipulates results when a specific visual trigger is detected.

Project Structure

• wrap.py: The main entry point for the attack. It wraps a trained model instance with the backdoor logic.

• attack/: Contains the core logic for backdoor injection, including task-type detection and target settings.

• validate/: Contains scripts to evaluate the performance of the poisoned model, comparing Clean Accuracy vs. Attack Success Rate (ASR).

• backdoor\_module.py: The implementation of the Backdoored(nn.Module) wrapper class.

---

Workflow

1. Attack Execution

To attack a model, you must run the wrap.py script. This script targets a **fully trained model instance** (e.g., a .pth file).

1. Loads the target model and extracts its state\_dict.

2. Automatically detects the task type (Classification, Regression, etc.) by inspecting the output layers.

3. Injects the Backdoored proxy, which monitors for a specific **3x3 checkerboard trigger**.

4. Saves the backdoored model instance to the specified path.

python wrap.py




```python

Attack demonstration video

demo.mp4