计算机科学
可靠性(半导体)
人工神经网络
可靠性工程
可靠性理论
可扩展性
人工智能
比例(比率)
深层神经网络
数据挖掘
机器学习
概率分布
深度学习
维修工程
反向传播
冗余(工程)
作者
Samira Nazari,Ali Azarpeyvand,Mohsen Afsharchi
标识
DOI:10.1109/tr.2026.3704288
摘要
Deep neural networks (DNNs) are increasingly deployed in reliability-critical applications. However, their susceptibility to hardware faults and runtime perturbations raises significant concerns about model-level dependability. This article presents GRADE, a scalable and model-agnostic framework for fine-grained reliability assessment and fault sensitivity analysis across diverse DNN architectures. GRADE computes hierarchical sensitivity scores (grades) at bit, weight, neuron, layer, and model levels, enabling rapid identification of critical vulnerabilities. Compared to both exhaustive and statistical fault injection techniques, GRADE achieves up to$6.4 \times 10^{6}$theoretical speedup and over$20 \times$faster runtime in practical settings, while maintaining comparable localization accuracy. The framework is evaluated on standard and well-controlled datasets (CIFAR-10, CIFAR-100, GTSRB, and Imagenet) and validated across vision transformers (ViTs), convolutional neural networks (CNNs), NLP models, and LLMs, demonstrating strong generalization and architecture-specific vulnerability patterns. The identified failure-prone components across different abstraction levels provide actionable insights for model hardening through targeted retraining, redundancy, and pruning strategies. By offering interpretable, quantitative, and efficient reliability metrics, GRADE contributes to the dependability engineering of AI systems and supports future integration into system-level reliability frameworks.
科研通智能强力驱动
Strongly Powered by AbleSci AI