SDEFL: A Lightweight Fault Detection and Localization Method for Deep Neural Networks*
Bo Yang, Jiawei Hu, Jialun Cao · 2024
Fault detection and localization in deep neural networks (DNN) refers to identifying and diagnosing the causes of errors or performance degradation in the learning process of the network. Faults include identifying incorrect weights, biases, activation functions, or network structures, which can be caused by problems such as overfitting, underfitting, vanishing gradients, or explosions. Fault detection and localization of DNN is a key task to ensure the performance, safety, and reliability of the model, which has important research value and application prospects for promoting the application and development of deep learning technology. The existing research on fault detection and localization of DNN includes rule-based methods and learningbased methods, which monitor the training process of deep learning models from multiple perspectives and locate the faults generated by the models when abnormal behaviors are found. However, these methods are carried out around the features of model failures, and lack of the code structure and syntax information of the model. In this paper, we propose a fault detection and localization method (SDEFL) based on the static structural fault features, dynamic training fault features, and program source code features based on AST of the deep neural network, which takes the static structure information of the neural network, the dynamic training information, and the syntax and semantic information of the representation program as the model source code as features, learns the relationship between the DNN and its fault class, and conducts code-level localization after identifying the fault class. A series of experiments have been conducted to evaluate SDEFL on data containing 48 type faults. The experimental results show that SDEFL delivers higher fault detection localization than the state-of-the-art techniques.