Jun Zhao

Kejian Zhu, Shangqing Tu, Zhuoran Jin et al.

The development of large language models (LLMs) depends on trustworthy evaluation. However, most current evaluations rely on public benchmarks, which are prone to data contamination issues that significantly compromise fairness. Previous researches have focused on constructing dynamic benchmarks to address contamination. However, continuously building new benchmarks is costly and cyclical. In this work, we aim to tackle contamination by analyzing the mechanisms of contaminated models themselves. Through our experiments, we discover that the overestimation of contaminated models is likely due to parameters acquiring shortcut solutions in training. We further propose a novel method for identifying shortcut neurons through comparative and causal analysis. Building on this, we introduce an evaluation method called shortcut neuron patching to suppress shortcut neurons. Experiments validate the effectiveness of our approach in mitigating contamination. Additionally, our evaluation results exhibit a strong linear correlation with MixEval, a recently released trustworthy benchmark, achieving a Spearman coefficient ($ρ$) exceeding 0.95. This high correlation indicates that our method closely reveals true capabilities of the models and is trustworthy. We conduct further experiments to demonstrate the generalizability of our method across various benchmarks and hyperparameter settings. Code: https://github.com/GaryStack/Trustworthy-Evaluation

Renyang Liu, Kwok-Yan Lam, Wei Zhou et al.

Many attack techniques have been proposed to explore the vulnerability of DNNs and further help to improve their robustness. Despite the significant progress made recently, existing black-box attack methods still suffer from unsatisfactory performance due to the vast number of queries needed to optimize desired perturbations. Besides, the other critical challenge is that adversarial examples built in a noise-adding manner are abnormal and struggle to successfully attack robust models, whose robustness is enhanced by adversarial training against small perturbations. There is no doubt that these two issues mentioned above will significantly increase the risk of exposure and result in a failure to dig deeply into the vulnerability of DNNs. Hence, it is necessary to evaluate DNNs' fragility sufficiently under query-limited settings in a non-additional way. In this paper, we propose the Spatial Transform Black-box Attack (STBA), a novel framework to craft formidable adversarial examples in the query-limited scenario. Specifically, STBA introduces a flow field to the high-frequency part of clean images to generate adversarial examples and adopts the following two processes to enhance their naturalness and significantly improve the query efficiency: a) we apply an estimated flow field to the high-frequency part of clean images to generate adversarial examples instead of introducing external noise to the benign image, and b) we leverage an efficient gradient estimation method based on a batch of samples to optimize such an ideal flow field under query-limited settings. Compared to existing score-based black-box baselines, extensive experiments indicated that STBA could effectively improve the imperceptibility of the adversarial examples and remarkably boost the attack success rate under query-limited settings.