深度神经网络在高铁运行环境识别中的鲁棒性验证

Robustness Verification of Deep Neural Networks on High-speed Rail Operating Environment Recognition

摘要:

改进了DeepTRE的实现，在保留DeepTRE验证能力的前提下大幅降低了DeepTRE的空间复杂度，以适应大规模数据集场景。在高铁运行环境识别场景中评估了改进后的DeepTRE，并与其他主流验证工具DLV和SafeCV对比。实验结果表明，改进后的DeepTRE工具的显存占用显著低于原DeepTRE工具，相较于其他神经网络验证工具，改进后的DeepTRE工具在具有较快验证速度的前提下拥有更优异的验证效果。

Abstract:

The implementation of DeepTRE was improved to adapt to large-scale dataset scenarios， which greatly reduces the space complexity of DeepTRE when retaining the excellent verification ability of DeepTRE. The improved DeepTRE was evaluated in the high-speed rail operating environment recognition scenarios and was compared with other mainstream verification tools， i.e.，DLV and SafeCV. The experimental results show that the memory usage of the improved DeepTRE tool is significantly lower than that of the original DeepTRE tool. Compared with other neural network verification tools， the improved DeepTRE tool has better verification effect on the premise of faster verification speed.