TerGEC : A Graph Enhanced Contrastive Approach for Program Termination Analysis
Research output: Journal Publications and Reviews › RGC 21 - Publication in refereed journal › peer-review
Author(s)
Related Research Unit(s)
Detail(s)
| Original language | English |
|---|---|
| Article number | 103141 |
| Journal / Publication | Science of Computer Programming |
| Online published | 15 May 2024 |
| Publication status | Online published - 15 May 2024 |
Link(s)
Abstract
Context : Programs with non-termination behavior induce various bugs, such as denial-of-service vulnerability and memory exhaustion. Hence the ability to detect non-termination programs before software deployment is crucial. Existing detection methods are either execution-based or deep learning-based. Despite great advances, their limitations are evident. The former requires complex sandbox environments for execution, while the latter lacks fine-grained analysis.
Objective : To overcome the above limitations, this paper proposes a graph-enhanced contrastive approach, namely TerGEC, which combines both inter-class and intra-class semantics to carry out a more fine-grained analysis and exempt execution during the detection process.
Methods : In detail, TerGEC analyzes behaviors of programs from Abstract Syntax Trees (ASTs), thereby capturing intra-class semantics both syntactically and lexically. Besides, it incorporates contrastive learning to learn the discrepancy between program behaviors of termination and non-termination, thereby acquiring inter-class semantics. In addition, graph augmentation is designed to improve the robustness. Weighted contrastive loss and focal loss are also equipped in TerGEC to alleviate the classes-imbalance problem during the non-termination detection. Consequently, the whole detection process can be handled more fine-grained, and the execution can also be exempted due to the nature of deep learning.
Results : We evaluate TerGEC on five datasets of both Python and C languages. Extensive experiments demonstrate TerGEC achieves the best performance overall. Among all experimented datasets, TerGEC outperforms state-of-the-art baselines by 8.20% in terms of mAP and by 17.07% in terms of AUC on average.
Conclusion : TerGEC is capable of detecting non-terminating programs with high precision, showing that the combination of inter-class and intra-class learning, along with our proposed classes-imbalance solutions, is significantly effective in practice.
Objective : To overcome the above limitations, this paper proposes a graph-enhanced contrastive approach, namely TerGEC, which combines both inter-class and intra-class semantics to carry out a more fine-grained analysis and exempt execution during the detection process.
Methods : In detail, TerGEC analyzes behaviors of programs from Abstract Syntax Trees (ASTs), thereby capturing intra-class semantics both syntactically and lexically. Besides, it incorporates contrastive learning to learn the discrepancy between program behaviors of termination and non-termination, thereby acquiring inter-class semantics. In addition, graph augmentation is designed to improve the robustness. Weighted contrastive loss and focal loss are also equipped in TerGEC to alleviate the classes-imbalance problem during the non-termination detection. Consequently, the whole detection process can be handled more fine-grained, and the execution can also be exempted due to the nature of deep learning.
Results : We evaluate TerGEC on five datasets of both Python and C languages. Extensive experiments demonstrate TerGEC achieves the best performance overall. Among all experimented datasets, TerGEC outperforms state-of-the-art baselines by 8.20% in terms of mAP and by 17.07% in terms of AUC on average.
Conclusion : TerGEC is capable of detecting non-terminating programs with high precision, showing that the combination of inter-class and intra-class learning, along with our proposed classes-imbalance solutions, is significantly effective in practice.
Research Area(s)
- Code Representation Learning, Graph Neural Networks, Contrastive Learning, Termination Analysis
Citation Format(s)
TerGEC: A Graph Enhanced Contrastive Approach for Program Termination Analysis. / Liu, Shuo; Keung, Jacky Wai; Yang, Zhen et al.
In: Science of Computer Programming, 15.05.2024.
In: Science of Computer Programming, 15.05.2024.
Research output: Journal Publications and Reviews › RGC 21 - Publication in refereed journal › peer-review
