The increasing use of deep neural networks (DNNs) in safety-critical systems has raised concerns about their potential for exhibiting ill-behaviors. While DNN verification and testing provide post hoc conclusions regarding unexpected behaviors, they do not prevent the erroneous behaviors from occurring. To address this issue, DNN repair/patch aims to eliminate unexpected predictions generated by defective DNNs. Two typical DNN repair paradigms are retraining and fine-tuning. However, existing methods focus on the high-level abstract interpretation or inference of state spaces, ignoring the underlying neurons' outputs. This renders patch processes computationally prohibitive and limited to piecewise linear (PWL) activation functions to great extent. To address these shortcomings, we propose a behavior-imitation based repair framework, BIRDNN, which integrates the two repair paradigms for the first time. BIRDNN corrects incorrect predictions of negative samples by imitating the closest expected behaviors of positive samples during the retraining repair procedure. For the fine-tuning repair process, BIRDNN analyzes the behavior differences of neurons on positive and negative samples to identify the most responsible neurons for the erroneous behaviors. To tackle more challenging domain-wise repair problems (DRPs), we synthesize BIRDNN with a domain behavior characterization technique to repair buggy DNNs in a probably approximated correct style. We also implement a prototype tool based on BIRDNN and evaluate it on ACAS Xu DNNs. Our experimental results show that BIRDNN can successfully repair buggy DNNs with significantly higher efficiency than state-of-the-art repair tools. Additionally, BIRDNN is highly compatible with different activation functions.
Credit assignment problem of neural networks refers to evaluating the credit of each network component to the final outputs. For an untrained neural network, approaches to tackling it have made great contributions to parameter update and model revolution during the training phase. This problem on trained neural networks receives rare attention, nevertheless, it plays an increasingly important role in neural network patch, specification and verification. Based on Koopman operator theory, this paper presents an alternative perspective of linear dynamics on dealing with the credit assignment problem for trained neural networks. Regarding a neural network as the composition of sub-dynamics series, we utilize step-delay embedding to capture snapshots of each component, characterizing the established mapping as exactly as possible. To circumvent the dimension-difference problem encountered during the embedding, a composition and decomposition of an auxiliary linear layer, termed minimal linear dimension alignment, is carefully designed with rigorous formal guarantee. Afterwards, each component is approximated by a Koopman operator and we derive the Jacobian matrix and its corresponding determinant, similar to backward propagation. Then, we can define a metric with algebraic interpretability for the credit assignment of each network component. Moreover, experiments conducted on typical neural networks demonstrate the effectiveness of the proposed method.