Proposal of Efficiency Metric for White-Box Deep Learning Testing

초록

Recent advancements in artificial intelligence (AI) have led to increased integration of AI technologies in various aspects of our lives. Deep learning systems, in particular, have proven effective in many areas. However, ensuring the quality, reliability, and safety of deep learning systems in critical environments is challenging. Traditional software testing methods are not suitable for deep learning systems. While there have been efforts to develop reliable testing techniques, there is a lack of focus on efficiency. As deep learning systems become larger and more complex, testing becomes more difficult and resource-intensive. The research community lacks a standardized efficiency metric for deep learning testing methods. We propose an efficiency metric that combines neuron coverage and test data accuracy to objectively evaluate the efficiency of white-box deep learning testing methods.