SDC vulnerability prediction method based on instruction feature importance

Abstract

The invention discloses an SDC vulnerability prediction method based on instruction feature importance. The method comprises the following steps: performing instruction feature extraction on a complex program; random fault injection is carried out on the complex program, and an instruction SDC vulnerability value is obtained; measuring the importance degree of the instruction SDC vulnerability feature T to an SDC vulnerability value based on a Gini coefficient, and constructing an instruction feature importance evaluation coefficient; the instruction feature importance is fused into the discourse right of each sub-decision tree of LightGBM, and a new discourse right omega is generated to improve the prediction precision of the instruction SDC vulnerability. Based on an improved LightGBM algorithm, constructing and training an SDC vulnerability prediction model based on instruction feature importance; and performing instruction SDC vulnerability prediction on the to-be-predicted program by the obtained final SDC vulnerability prediction model. Compared with other methods, the method has higher vulnerability prediction precision, can effectively reduce memory consumption and calculation cost, can be applied to standard data types and mixed data types, and has the advantages of light weight, high interpretation and the like.

Description

technical field [0001] The invention belongs to the field of software detection and reinforcement, reliability and security, in particular to an SDC vulnerability prediction method, system, computer equipment and storage medium based on the importance of instruction features. Background technique [0002] With the rapid development of 5G communication, the integration of transistors in the chip manufacturing process is getting higher and higher. These chips are widely used in industrial automation, medical services, smart home and other fields. However, soft errors (e.g. signal or data errors) are likely to result due to huge resource consumption (e.g. CPU cores and memory) in various devices. Among them, silent data corruption (SDC) is the most dangerous error among soft errors, which manifests as a bit flip phenomenon in multi-control, distributed storage, and public cloud storage but cannot be caught by the system. Therefore, establishing an SDC detection model accurate...

AI Technical Summary

Problems solved by technology

However, for SDC detection, a large number of fault injections brings a huge time cost

Images