Distributed Repair of Deep Neural Networks (Hot off the Press at GECCO 2024)

Davide Li Calsi, Matias Duran, Xiao-Yi Zhang, Paolo Arcaini, Fuyuki Ishikawa

Abstract: Deep Neural Networks (DNNs) are increasingly used for critical tasks, such as classification in autonomous driving, whose trustworthiness is extremely important. To guarantee trustworthiness, DNN repair has been used to improve DNN performance, by using metaheuristic search to find alternative values of specific weights, that allow to improve the DNN accuracy. However, achieving perfect accuracy is not possible, and, therefore, one should prioritise the most critical misclassifications, such as those of pedestrians. To this aim, we propose DistrRep, a search-based DNN repair approach that considers priorities among the different misclassifications given by their risk levels. For each misclassification, DistrRep identifies the weights responsible for that misclassification, and runs a repair approach based on Particle Swarm Optimization that fixes the weights. Then, starting from all the repaired models, it runs another search-based repair that searches for the DNN model that integrates all the single repaired models, by considering the risk levels of the different misclassifications. Experimental results show that the search-based approach implemented by DistrRep is more effective that retraining approaches and other DNN repair approaches.This is an extended abstract of the paper: D. Li Calsi, M. Duran, X. Zhang, P. Arcaini, and F. Ishikawa, "Distributed Repair of Deep Neural Networks", in 16th IEEE Conference on Software Testing, Verification and Validation (ICST 2023).