基于图神经网络的程序脆弱性指数评估方法

软错误会导致隐性偏差，严重影响计算机系统的可靠性。计算程序脆弱性指数是防护隐性偏差的先决条件。针对传统方法中程序语义提取不足，无法全面反映错误传播机理的问题，提出了一种基于图注意力网络的程序脆弱性指数评估方法EpicGNN。将脆弱性指数预测的任务转换为图神经网络的图回归任务，应用不同类型的边来表示不同的指令关系；引入结构化多头自注意力机制量化节点间、节点到图在错误传播中的重要程度；依据该重要性聚合节点信息、图信息形成图的表示向量，并利用回归模型预测脆弱性指数。实验结果表明，EpicGNN在spec2000、spec2006、rodinia等数据集上的平均绝对误差相比现有模型减少了0.037～0.258,对未见过的图仍然有良好的泛化性能。

Evaluation of program vulnerability factor based on graph neural network

Soft error can lead to silent data corruption(SDC) that affects the reliability of computer system. One prerequisite to prevent SDC is calculating the vulnerability factor of the target program. Traditional methods were not capable of extracting program semantics, which led to inferior of the fault propagation mechanism. This paper proposed a program vulnerability factor evaluation method based on graph attention network(EpicGNN). EpicGNN used structural multi-head self-attention to quantify importance of fault propagations from one node to its neighbors and further to graph, different types of edges represent different instruction relationships. Then it aggregated the information of node and graph to update the representation. It applied a regression model to predict vulnerability factor. Experimental on spec2000, spec2006, rodinia and other datasets achieve 0.037~0.258 lower average absolute error compared with traditional methods. Moreover, EpicGNN obtains good performance on unseen graphs.