Abstract:In safety-critical systems (e.g., autonomous vehicles and robots), Deep Neural Networks (DNNs) are becoming a key component for computer vision tasks, particularly semantic segmentation. Further, since the DNN behavior cannot be assessed through code inspection and analysis, test automation has become an essential activity to gain confidence in the reliability of DNNs. Unfortunately, state-of-the-art automated testing solutions largely rely on simulators, whose fidelity is always imperfect, thus affecting the validity of test results. To address such limitations, we propose to combine meta-heuristic search, used to explore the input space using simulators, with Generative Adversarial Networks (GANs), to transform the data generated by simulators into realistic input images. Such images can be used both to assess the DNN performance and to retrain the DNN more effectively. We applied our approach to a state-of-the-art DNN performing semantic segmentation and demonstrated that it outperforms a state-of-the-art GAN-based testing solution and several baselines. Specifically, it leads to the largest number of diverse images leading to the worst DNN performance. Further, the images generated with our approach, lead to the highest improvement in DNN performance when used for retraining. In conclusion, we suggest to always integrate GAN components when performing search-driven, simulator-based testing.

What problem does this paper attempt to address?

The key problem that this paper attempts to solve is the reliability testing and retraining of deep neural networks (DNNs) in safety - critical systems, such as self - driving cars and robots. Specifically, the author focuses on how to improve the reliability and performance of DNNs in computer vision tasks by combining meta - heuristic search, simulators, and generative adversarial networks (GANs). ### Problem Background 1. **Applications of DNNs in Safety - Critical Systems** - DNNs are becoming increasingly important in computer vision tasks (such as semantic segmentation). - Since the behavior of DNNs cannot be evaluated through code inspection and analysis, automated testing has become an important means to ensure their reliability. 2. **Limitations of Existing Testing Methods** - Existing automated testing solutions mainly rely on simulators, and the fidelity of simulators is always limited, which affects the effectiveness of test results. - The images generated by simulators may not be consistent with the real - world distribution, resulting in poor performance of DNNs in practical applications. ### The Method Proposed in the Paper To overcome the above limitations, the author proposes a method named DESIGNATE, which combines the following techniques: - **Meta - heuristic Search**: Used to explore the input space and generate diverse inputs that may cause DNNs to fail. - **Simulators**: Generate initial simulated images and corresponding ground - truth information (such as segmentation maps). - **Generative Adversarial Networks (GANs)**: Convert the images generated by simulators into realistic images to better reflect the real - world data distribution. ### Specific Objectives 1. **Improve the Effectiveness of Testing** - Identify potential problems of DNNs by generating more diverse and closer - to - real - world images. 2. **Improve the Performance of DNNs** - Retrain DNNs using the generated realistic images, thereby improving their reliability in practical applications. ### Experimental Verification The author has proven through experiments that DESIGNATE is superior to existing testing methods in the following aspects: - **Generate More Diverse Failure Scenarios**: DESIGNATE can generate more different scenarios that will cause a decline in DNN performance. - **Improve the Retraining Effect of DNNs**: After retraining with the images generated by DESIGNATE, the accuracy of DNNs has increased by 8 percentage points. ### Conclusion The author suggests that the GAN component should always be integrated in simulator - based DNN testing to ensure the representativeness and effectiveness of test results and improve the reliability of DNNs in practical applications. --- In summary, this paper aims to solve the problems of reliability testing and retraining of DNNs in safety - critical systems by combining meta - heuristic search, simulators, and GANs, thereby ensuring the performance and reliability of DNNs in practical applications.

Search-based DNN Testing and Retraining with GAN-enhanced Simulations

See Clearer at Night: Towards Robust Nighttime Semantic Segmentation Through Day-Night Image Conversion

SIMGAN: Photo-Realistic Semantic Image Manipulation Using Generative Adversarial Networks.

A Search-Based Testing Framework for Deep Neural Networks of Source Code Embedding

DeepRoad: GAN-based Metamorphic Autonomous Driving System Testing

Application of GANs-based virtual environment generation in automatic driving simulation training

DiffGAN: A Test Generation Approach for Differential Testing of Deep Neural Networks

Simulator-based explanation and debugging of hazard-triggering events in DNN-based safety-critical systems

Automated Testing for Deep Learning Systems with Differential Behavior Criteria

Toward Efficiently Evaluating the Robustness of Deep Neural Networks in IoT Systems: A GAN-Based Method

DeepEvolution: A Search-Based Testing Approach for Deep Neural Networks

Coarse-to-Fine Searching for Efficient Generative Adversarial Networks.

Survey on Generative Adversarial Behavior in Artificial Neural Tasks

SADA: Semantic Adversarial Diagnostic Attacks for Autonomous Applications

Black-box Safety Analysis and Retraining of DNNs based on Feature Extraction and Clustering

Generative Model-Based Test Case Generation and Operational Testing for Deep Learning

Data Generation and Evaluation Using Deep Learning

EAGAN: Efficient Two-stage Evolutionary Architecture Search for GANs

Data-driven Verification of DNNs for Object Recognition

SilGAN: Generating driving maneuvers for scenario-based software-in-the-loop testing

A Deeply Supervised Semantic Segmentation Method Based on GAN