Xiang Aoling^*

School of Computer Science, Guangdong University of Technology

Abstract：

This paper presents a preliminary study on the impact of quantization on the robustness of convolutional neural networks (CNNs). Quantization is an important technique for reducing model size and computational cost, but low-bit representation may introduce rounding errors, clipping effects and activation distribution shifts. These changes can weaken model stability when the input contains natural noise or adversarial perturbations. Instead of reporting unsupported experimental conclusions, this draft constructs a clear evaluation framework for comparing full-precision CNNs with post-training quantization and quantization-aware training under multiple robustness conditions. The proposed mitigation framework combines calibration-aware clipping, layer-wise bit-width selection and robustness-oriented fine-tuning. The purpose of this study is to clarify the possible mechanism of robustness degradation and provide an executable experimental plan for later empirical verification. This work can support the deployment of efficient and reliable CNN models on resource-constrained devices.

Key Words：

convolutional neural network; quantization; robustness; adversarial perturbation; model compression