Efficient and Systematic Metamorphic Testing for Vision-Based Autonomous Driving

This repository provides the main implementation of the Efficient and Systematic Metamorphic Testing Framework for Vision-Based Autonomous Driving Systems (ADS). The framework integrates causal inference and diffusion models to enhance the reliability testing of ADS, facilitating the detection of faults under diverse driving conditions.

Please Kindly NOTE:

The code is currently undergoing necessary refinement and will be available soon.

Folder Structure

• Autopilot/: Contains 5 implementations of the tested autonomous driving algorithms and the original test images that need to be mutated

• CausalEffect/: Structural causal models, counterfactual models, and a multi-objective search framework for critical test conditions.

• Metamorphic/: Apply the test conditions found in the semantic description, the low-risk test images are generalized to the high-risk driving conditions based on the fine-tuned diffusion model.

• TACTIC/: Representative Metamorphic Testing baselines for comparison.

  • do_testing/: Main Implementations to perform TACTIC test
  • logger/: TACTIC search records
  • munit/: Contains random MUNIT method and DeepRoad method
  • test_outputs/: Contains the mutation results of the baselines
  • train_outputs/: Neuron coverage during training of the algorithm under test

• requirements.txt: A list of Python dependencies required to run the project.