BIRDNN: Behavior-Imitation Based Repair for Deep Neural Networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Deep neural networks (DNNs) are increasingly used in safety-critical systems.
• Concerns exist regarding DNNs exhibiting undesirable behaviors and making incorrect predictions.
• Existing DNN repair methods are computationally intensive and limited in scope.

Purpose of the Study:

• To introduce BIRDNN, a behavior-imitation based DNN repair framework.
• To overcome limitations of existing retraining- and fine-tuning-based DNN repair methods.
• To address domain-wise repair problems (DRPs) more effectively.

Main Methods:

• BIRDNN employs a sampling technique to characterize DNN domain behaviors.
• For retraining, it imitates expected positive sample behaviors to rectify incorrect predictions.
• For fine-tuning, it analyzes neuron behavior differences and uses particle swarm optimization (PSO).

Main Results:

• BIRDNN was evaluated on ACAS Xu and MNIST DRP benchmarks.
• The framework demonstrated superior effectiveness, efficiency, and compatibility in repairing DNNs.
• Experimental results show comprehensive repair capabilities compared to state-of-the-art methods.

Conclusions:

• BIRDNN offers a robust and versatile solution for DNN repair.
• The behavior-imitation approach addresses limitations of current patching strategies.
• BIRDNN provides a computationally efficient and broadly applicable DNN repair framework.