MetaPri: Test Input Prioritization Based on Metamorphic Testing

Owing to the extensive application of deep neural networks (DNNs) in safety-critical tasks, it is necessary to thoroughly test DNN models before model deployment to mitigate risks. However, this typically requires expensive manual labeling and validation, with test cases inputting DNNs to discover erroneous decisions. To address the issues of testing efficiency and testing label costs, previous research has used a prioritization testing approach based on neuron coverage, but such a method proves ineffective in the context of test case prioritization. Other researchers have also attempted mutation-based approaches, but this method is associated with high time costs and low efficiency. To address these shortcomings, in this paper, we propose a DNNs test input prioritization method based on a metamorphosis test called MetaPri. By utilizing a novel diversity metric based on the internal neuron gradients of DNNs, MetaPri mitigates the cost of manual labeling and detects violations. We evaluate the proposed method based on effectiveness, guidance. Experimental results show that MetaPri improves the model accuracy by an average of 9.92% under different confrontation scenarios, which significantly outperforms other baseline methods, as well as verifies its versatility and effectiveness.