GNTF: A Lightweight CNN Robustness Enhancement Method for IoT Devices

Deploying lightweight convolutional neural networks (CNNs) to provide vision services on resource-constrained Internet of Things (IoT) devices has become the mainstream approach to addressing computing and energy consumption constraints. However, these IoT devices often operate in complex outdoor environments (e.g., fog, rain, and snow), and the quality of the data they collect is easily degraded, causing standard lightweight CNNs to experience a significant performance drop under such corrupted data. To this end, this paper proposes a Generative Nonlinear Transformation Filter (GNTF) method to improve the generalization performance of lightweight CNNs on corrupted data. The core of the GNTF is that only a portion of the filters are used as learnable parameters (named seed filters), while the remaining filters are generated by applying the nonlinear transformation to the seed filters, which is randomly initialized and fixed during training. This design makes the model parameters less dependent on the training data distribution, thereby regularizing the model, mitigating overfitting, and enhancing its robustness to data degradation. The GNTF further analyzes the structural characteristics of lightweight CNNs, showing that significant performance improvements can be achieved simply by replacing the depthwise convolutional modules. Furthermore, this paper examines the properties of various nonlinear transformation functions and finds that model robustness can be improved by applying simple translations. To verify the effectiveness of the GNTF, we conducted extensive experiments on the CIFAR-10/-100, CIFAR-10-C/-100-C, and ICONS-50 datasets, using the MobileNetV2, ShuffleNetV2, EfficientNet, and GhostNet models. The results show that the proposed GNTF can improve the model's accuracy on corrupted data while reducing the number of trainable parameters in most cases. For example, on the CIFAR-10-C dataset, ShuffleNetV2 with the GNTF improves accuracy by about 3.3% over the original model while slightly reducing the number of trainable parameters.