Saihua Cai

Context: Verifying the safety and robustness of autonomous driving systems is a core challenge for commercialization. Existing testing methods have complementary shortcomings: fuzzing tests efficiently explore the input space but lack semantic oracles for output assessment, whereas metamorphic tests circumvent the oracle problem yet rely on manually defined relations, resulting in limited scenario generation efficiency.Objective: We propose DSFuzz, a robustness verification framework that integrates fuzzing and metamorphic testing to synergistically address these limitations, providing both semantic guidance and automated oracles for autonomous driving systems.Method: DSFuzz adopts fuzzing as the overall test driver and incorporates CGAN and VAE‐GAN to perform semantically guided seed mutation, generating diverse and realistic test scenarios while preserving semantic plausibility. It further defines three metamorphic relations along the dimensions of perceptual tolerance, semantic robustness, and output consistency, offering interpretable and automated evaluation criteria for model behavior.Results: Experiments on the CARLA simulation platform with four typical driving models (Dave‐orig, Dave‐norminit, Dave‐dropout, and Epoch) show that DSFuzz significantly outperforms baseline fuzzing methods in error trigger rate. The Simpson’s Diversity Index confirms the diversity and balanced distribution of the generated scenarios. Moreover, retraining the models on error images produced by DSFuzz reduces the error rate by up to 21.6% in complex scenarios, demonstrating its dual value in defect detection and model enhancement.Conclusion: DSFuzz provides an efficient, scalable, and semantically rich solution for robustness verification of autonomous driving systems, effectively bridging the gap between exploration efficiency and semantic validity in safety testing.

The detection of software vulnerabilities is an essential security issue in the field of cyberspace security. However, existing detectors often suffer from two main limitations: low computational efficiency owing to high cost and poor scalability when processing large‐scale code, and inadequate feature representation capability due to their failure to capture sufficient semantic and structural properties of programs, which ultimately leads to decreased detection accuracy. To address these challenges and improve detection accuracy, we propose MCV‐DCNN, a novel software vulnerability detection method based on multi‐scale centrality‐weighted image generation and dual‐channel convolutional neural networks. The proposed MCV‐DCNN method leverages multiple node centrality metrics to transform source code into image representations, thereby capturing a more holistic view of the code's structural properties. It also employs a dual‐channel convolutional neural network architecture to independently process one image with each channel in parallel. This design mitigates information interference and redundancy between the representations during feature learning. We evaluate MCV‐DCNN method on a public dataset with 33,362 C/C++ programs. Experimental results demonstrate that it significantly improves vulnerability detection accuracy while maintaining efficiency. Furthermore, the evaluation on 26,193 real‐world functions shows that MCV‐DCNN outperforms baselines by detecting 505 additional vulnerabilities, validating its effectiveness in practical settings.

With the rapid growth of network traffic, data has become increasingly complex and voluminous, posing significant challenges for accurate anomaly detection. Traditional deep learning approaches often fail to capture the rich interdependencies and heterogeneous nature of traffic features, limiting their effectiveness in identifying subtle or evolving abnormal patterns. To address these challenges, this paper proposes GS-HF (GraphSAGE with heterogeneous features), an anomaly detection framework. The method integrates both statistical and image-like features extracted from raw traffic data to capture multi-dimensional characteristics, and then constructs a graph based on the fused heterogeneous features to better represent relationships among traffic flows. An improved GraphSAGE model is applied for detection, incorporating Focal Loss to handle class imbalance in real-world anomaly datasets. Extensive experiments on multiple network traffic datasets demonstrate that GS-HF achieves superior detection performance compared to existing methods, highlighting its robustness and effectiveness in handling diverse and complex traffic patterns.

With the rapid development of deep learning–based autonomous driving technology, how to verify and improve the robustness and safety of autonomous driving models and systems has become a research hotspot in academia and industry. The environment faced by autonomous driving systems is diverse, dynamic, and complex, which makes its testing work face many problems such as high‐dimensional input data, environmental uncertainty, and black‐box characteristics of models. Therefore, this paper systematically sorts out the current testing methods for autonomous driving models and systems, focusing on two mainstream testing methods: metamorphic testing and fuzzing testing. In terms of metamorphic testing, this paper deeply analyzes the design challenges of Test Oracle and explores how to effectively transform complex scenarios to ensure the comprehensiveness and accuracy of the test. In terms of fuzzing testing, this paper focuses on the application of optimization search technology in generating test scenarios, especially the effects of technologies such as genetic algorithms and neural networks in discovering model failure points and improving system robustness. Furthermore, this paper also discusses how to effectively combine these two mainstream testing methods, summarizes and generalizes the commonly used datasets, models and simulation simulators in these two tests, and also discusses the key points and improvement suggestions of each method. It strives to fully present the application status and development trend of current testing technology. A comparative analysis of representative methods is conducted using metrics like violation rate and unique violations, revealing performance trends under unified conditions. Finally, this paper points out the limitations of current research and proposes possible directions for future research. This article aims to provide reference and guidance for in‐depth research on autonomous driving model testing technology and promote the further development of autonomous driving systems in terms of safety and reliability.