Antidote or Placebo? Unraveling the Efficacy of Neuron Coverage Criteria on Testing Transformer-based Language Models

Abstract

In the realm of deep learning, a variety of neuron coverage criteria for Deep Neural Networks (DNNs) have been devised to effectively assess the quality of test suites and facilitate the generation of test inputs. Recently proposed coverage criteria, incorporating representation distribution and causal relationships, have infused fresh vitality into this field. However, the focus of previous works is primarily on Convolutional Neural Networks for computer vision, leading to a research gap in exploring coverage testing for language models. Concurrently, with the rise of large language models, transformer-based language models have become increasingly dominant, and numerous ones have sprouted. Therefore, the effectiveness of coverage criteria in transformer-based language tasks, especially with the introduction of novel criteria, remains an unresolved open problem. To tackle it, this study examines these concerns by evaluating a wide range of criteria, including four well-established ones and two state-of-the-art criteria, across three types of transformer-based models: encoder-only, decoder-only, and encoder-decoder models. Building on previous research, we conduct a comprehensive evaluation across three key areas: regarding test suite properties, 1) Error-revealing capability, i.e., sensitivity to adversarial examples; 2) Diversity, i.e., distribution diversity and sample fairness (category diversity); and regarding test suite generation, 3) Input generation guidance, i.e., the ability to guide the generation of more valuable samples. The experimental results demonstrate that the impact of coverage criteria is multifaceted. For the error-revealing capability of test suites, the additional coverage for erroneous samples over noise samples is only 0.32%. In terms of distribution diversity and sample fairness, 26 and 30 cases, respectively, out of 33 configurations are effectively evaluated. Additionally, incorporating neuron-wise coverage guidance during test suite generation slightly increases the production of adversarial samples by 4.56%. In conclusion, while current coverage criteria can act as an antidote for assessing simple diversity, they remain largely a placebo for the core task of revealing adversarial errors, particularly when relying on individual criterion. Consequently, their practical application requires carefully evaluating the trade-off between computational overhead and potential benefits given the massive scale of Transformers. However, this low cost-effectiveness ultimately highlights the urgent need to explore and develop more robust and efficient criteria designed specifically for Transformer-based models.

Ähnliche Arbeiten