BIRDNN: Behavior-Imitation Based Repair for Deep Neural Networks

The increasing utilization of deep neural networks (DNNs) in safety-critical systems has raised concerns about their potential to exhibit undesirable behaviors. Consequently, DNN repair/patching arises in response to the times, and it aims to eliminate unexpected predictions generated by flawed DNNs. However, existing repair methods, both retraining- and fine-tuning-based, primarily focus on high-level abstract interpretations or inferences of state spaces, often neglecting the outputs of underlying neurons. As a result, present patching strategies become computationally prohibitive and own restricted application scope (often limited to DNNs with piecewise linear (PWL) activation functions), particularly for domain-wise repair problems (DRPs). To overcome these limitations, we introduce BIRDNN, a behavior-imitation based DNN repair framework that supports alternative retraining and fine-tuning repair paradigms for DRPs. BIRDNN employs a sampling technique to characterize DNN domain behaviors and rectifies incorrect predictions by imitating the expected behaviors of positive samples during the retraining-based repair process. As for the fine-tuning repair strategy, BIRDNN analyzes the behavior differences of neurons between positive and negative samples to pinpoint the most responsible neurons for erroneous behaviors, and then integrates particle swarm optimization algorithm (PSO) to fine-tune buggy DNNs locally. Furthermore, we have developed a prototype tool for BIRDNN and evaluated its performance on two widely used DRP benchmarks, the ACAS Xu DNN safety repair problem and the MNIST DNN robustness repair problem. The experiments demonstrate that BIRDNN features more excellent effectiveness, efficiency, and compatibility in repairing buggy DNNs comprehensively compared with state-of-the-art repair methods.