面向边缘部署场景的轻量神经网络修复算法  

Light-Weight Neural Network Repair for Edge Computing Scenarios

作　　者：方毓楚 李文中[1] 曾曜 郑阳 胡崝 陆桑璐[1] FANG Yu-Chu;LI Wen-Zhong;ZENG Yao;ZHENG Yang;HU Zheng;LU Sang-Lu(Department of Computer Science and Technology,Nanjing University,Nanjing 210023;Trustworthiness Theory,Technology&Engineering Lab(TTE Lab),Huawei Technologies Co,.Ltd.,Shenzhen,Guangdong 518129)

机构地区：[1]南京大学计算机科学与技术系,南京210023 [2]华为技术有限公司可信理论技术与工程实验室,广东深圳518129

出　　处：《计算机学报》2024年第6期1413-1430,共18页Chinese Journal of Computers

基　　金：国家自然科学基金面上项目(61972196);国家自然科学基金重点项目(61832008,61832005);江苏省前沿引领技术基础研究重大项目(BK20222003)资助.

摘　　要：随着深度学习技术的不断进步,神经网络在各领域得到广泛应用,特别是在边缘计算环境中,例如智能交通和新型电网等典型场景.然而,神经网络的可靠性问题限制了其在真实世界的广泛应用.在复杂的边缘环境中,预训练模型往往因未涵盖所有可能的边缘情况而性能下降.因此,针对部署中的神经网络进行高效修复成为一个关键的研究课题.传统修复方法通常涉及整个模型的重新训练,这在边缘场景中具有诸多局限性.首先,不同地理区域的设备可能面临独特的自然噪声,使得统一模型难以适应所有环境.其次,深度神经网络的大规模参数使得其训练和部署时资源消耗巨大,且更新期间的服务中断将降低系统的可用性.为解决这些问题,本文提出了一种轻量级的补丁式神经网络修复算法.该算法通过引入个性化的补丁来增强神经网络对不同边缘环境中自然噪声和边角案例的鲁棒性.具体的,在故障定位阶段,类比于程序插桩中通过注入代码以检测、改进和分析软件行为,本文提出了神经网络插桩技术.通过将模型探针插桩进神经网络,观测其内部运行情况,实现了对错误样本的故障定位.在故障修复时,通过插入无监督搜索得到的神经网络补丁来纠正原始神经网络的输出.此外,本文提出了故障预测模块以提前预测潜在的错误输出,从而仅在必要时激活补丁.在基于2个数据集、15种噪声以及4个神经网络模型的实验中,与现有修复算法相比,本文方法在修复性能上取得了 6.64%至20.00%的提升.同时,本文方法所需的训练样本量减少了超过90%,而所需更新的参数量最高减少了 91.94%.这种有效且轻量的特性为解决边缘计算环境中神经网络的可靠性问题提供了有效途径.The continuous evolution and progress of deep learning technology have ushered in an era where neural networks play a pivotal role in various fields.This is particularly evident in edge computing environments such as intelligent transportation systems and next-generation power grids.However,despite the widespread applications,the reliability of neural networks remains a significant bottleneck,restricting their full potential in real-world scenarios.One of the primary challenges arises in complex edge environments,where pre-trained models often suffer from performance degradation due to the inherent difficulty of covering all possible edge scenarios comprehensively.Consequently,the need for an efficient repair mechanism for deployed neural networks has become a paramount focus of research.Traditional methods of repairing neural networks typically involve the cumbersome process of retraining the entire model.This approach presents several challenges,especially in edge scenarios.Firstly,devices located in different geographical regions may encounter unique natural noise,making it a challenging task for a unified model to seamlessly adapt to all diverse environments.Secondly,the large-scale parameters of deep neural networks contribute to substantial resource consumption during training and deployment,and the potential for service interruptions during updates poses a threat to system availability.To address these challenges head-on,this paper proposes a lightweight patch-based neural network repair algorithm.The core objective of this algorithm is to augment the robustness of neural networks against natural noise and corner cases prevalent in different edge environments by introducing personalized patches.The fault localization phase is a key aspect of this algorithm,drawing inspiration from probes in program instrumentation used to detect,improve,and analyze software behavior.In the context of neural networks,the paper introduces neural network instrumentation technology.This involves strategically inserting model p

关 键 词：神经网络修复 深度边缘计算 故障定位 故障预测 神经网络补丁 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]