Enhancing software reliability mitigates failures and reduces maintenance costs. In the LLVM compiler, different option sequences applied to the intermediate representation (IR) produce binaries with varying reliability levels, making the search for an optimal sequence in reliability-oriented compilation a key challenge. Although reinforcement learning (RL) has been employed to automate this process, existing methods suffer from two major limitations: they rely on structure-based static IR embeddings that overlook the dynamic behaviors induced by compilation transformations, leading to low-fidelity state representations, and construct action spaces that fail to preserve critical inter-option dependencies while maintaining spatial compactness, thereby impairing training efficiency and constraining optimization gains. This article presents Optimizing Software Reliability at Compile-time using Reinforcement Learning (OSRC-RL), which integrates two novel components: a Compilation Behavior-Aware State Representation (CBA-SR) and an Action Space Construction via Option Dependency Awareness and space minimization (ASC-ODA). CBA-SR jointly encodes program semantics and compilation dynamics, using activated options as multi-label supervisory signals to guide an expressive Graph Isomorphism Network in learning IR subgraph patterns correlated with option activations, yielding high-fidelity states that enable more reliable policy learning and faster convergence. ASC-ODA reconciles dependency preservation with spatial compactness by constructing Option Dependency Graph, performing dependency-aware hierarchical clustering and recursive intra-cluster path analysis, and applying a preference-aware selection to generate a compact yet dependency-preserving action space that enhances sample efficiency and policy stability. Evaluations across 20 benchmarks show that OSRC-RL achieves the highest average reliability gain ( \(I_{rg}{=}0.6694\) ) and competitive convergence speed compared with five baselines. Ablations attribute these gains to its core components: CBA-SR (up to 11.34% improvement via IR-graph learning) and ASC-ODA (up to 15.22% improvement via dependency-faithful actions). On industrial instances, it attains a 0.1818 average gain, surpassing the next-best method by 33.77%. Bounded overheads are mitigated by Lightweight Post-Processing (LPP), ensuring practical feasibility.

Autonomous Transportation Systems (ATS) make the transportation system transition from “passive transportation” to “autonomous service”. The wireless nature of communication in ATS presents significant cybersecurity challenges. Conventional intelligent vehicular intrusion detection methods may not suffice in situations where vehicular data is produced at an unprecedented scale and diverse cybersecurity threats are launched. Therefore, there is a demand for the creation of advanced intelligent vehicular intrusion detection systems that can effectively manage potential cyberattacks within ATS. Toward this end, this paper proposes QnGSA (hypercube driven graph self-attention intelligent vehicular intrusion detection model) in ATS, a novel intrusion detection model that helps to protect both the vehicles and the data they transmit, preventing disruptions to services, theft of sensitive information, and potential harm to passengers or cargo. QnGSA not only proposes a construction method of association graph by introducing hypercube and semi-supervised K-means++ clustering algorithm (QnSSKM). But also, QnGSA self-extracts the graph structural information of hypercube, and uses the graph self-attention mechanism to aggregate node features and obtain more accurate representation. Furthermore, this paper uses Graph Attention Network classifier to correlate the learned node representation with the fault category, and uses Softmax function to map the node representation to the probability distribution of different categories. The category with the highest probability is selected as the prediction label of the node, so as to realize the intelligent vehicular intrusion detection. Experiments results show that our proposed QnGSA method achieves the best results compared with state-of-the-art methods in terms of accuracy, macro precision/recall/F1.

Neural networks are extensively employed in safety-critical systems. However, these critical systems incorporating neural networks continue to pose risks due to the presence of adversarial examples. Although the security of neural networks can be enhanced by verification, verifying neural networks is an NP-hard problem, making the application of verification algorithms to large-scale neural networks a challenging task. For this reason, we propose NNTBFV, a framework that utilizes the principles of Testing-Based Formal Verification (TBFV) to simplify neural networks and verify the simplified networks. Unlike conventional neural network pruning techniques, this approach is based on specifications, with the goal of deriving approximate execution paths under given preconditions. To mitigate the potential issue of unverifiable conditions due to overly broad preconditions, we also propose a precondition partition method. Empirical evidence shows that as the range of preconditions narrows, the size of the execution paths also reduces accordingly. The execution path generated by NNTBFV is still a neural network, so it can be verified by verification tools. In response to the results from the verification tool, we provide a theoretical method for analysis. We evaluate the effectiveness of NNTBFV on the ACAS Xu model project, choosing Verification-based and Random-based neural network simplification algorithms as the baselines for NNTBFV. Experiment results show that NNTBFV can effectively approximate the baseline in terms of simplification capability, and it surpasses the efficiency of the random-based method.

A deep learning system (DLS) developed based on one software project for defect prediction may well be applied to the related code on the same project but is usually difficult to be applied to new or unknown software projects. To address this problem, we propose a Transferable Graph Convolutional Neural Network (TGCNN) that can learn defects from the lightweight semantic graphs of code and transfer the learned knowledge from the source project to the target project. We discuss how the semantic graph is constructed from code; how the TGCNN can learn from the graph; and how the learned knowledge can be transferred to a new or unknown project. We also conduct a controlled experiment to evaluate our method. The result shows that despite some limitations, our method performs considerably better than existing methods.

Software defect prediction is aimed at helping developers to quickly locate defective components in the code repository and thus better allocate resources. However, most of the current traditional defect prediction methods mainly depend on the design of static metrics, but these methods ignore the semantic and structural information of the code. As a result, researchers have turned to building models by extracting semantic features from code through abstract syntax trees. In this article, we introducegated homogeneous fusion networkfor defect prediction namely GHFNet, jointing high-level semantic feature extraction and weighted static feature extraction. Through the mechanism of homogeneous gating fusion, weights are adaptively assigned to the two types of features based on the correlation of these features to form fused features for defect prediction in the code. Experimental results show that the proposed approach is a significant improvement. Specifically, for the reference method we present GHFNet improved from 9.4 to 15.2 percentage points in effort-unaware scenarios (F-measure) and from 3.7 to 8 percentage points in effort-aware scenarios (Popt) for defect prediction.

Software defect prediction plays a significant role in the software development cycle but suffers from many difficulties. In this paper we propose a novel deep learning model (including algorithms) called Extractor, Parser, and Reviewer (EPR) for defect prediction in software. Two different networks, recurrent neural networks (RNNs) and one-dimensional convolutional networks(ODCNs), are employed by the EPR for different purposes. RNN is utilized to extract contextual features to represent semantic dependencies between code tokens and ODCN acts as a parser to establish dependencies between semantic features. Meanwhile, the attention mechanism of the two networks is used as a reviewer to assign different weights from location information to the importance of the features, respectively. Our proposed model is validated by the PROMISE repository, and the results show that the proposed model in this paper significantly outperforms several existing algorithms.