GNTF: A Lightweight CNN Robustness Enhancement Method for IoT Devices

Area of Science:

• Computer Vision
• Machine Learning
• Internet of Things (IoT)

Background:

• Lightweight convolutional neural networks (CNNs) are crucial for vision tasks on resource-constrained Internet of Things (IoT) devices.
• Performance of these CNNs degrades significantly in outdoor environments due to corrupted data (e.g., fog, rain).

Purpose of the Study:

• To propose a Generative Nonlinear Transformation Filter (GNTF) method to enhance the generalization performance of lightweight CNNs on corrupted data.
• To improve the robustness and accuracy of CNNs deployed on IoT devices operating in adverse conditions.

Main Methods:

• The GNTF method utilizes a subset of filters as learnable parameters (seed filters).
• Remaining filters are generated via nonlinear transformation of seed filters, which are fixed during training.
• The method focuses on replacing depthwise convolutional modules and applying simple translation transformations.

Main Results:

• The GNTF method demonstrated improved accuracy on corrupted datasets (CIFAR-10-C, CIFAR-100-C) across various lightweight CNN architectures.
• Experiments showed enhanced robustness and generalization performance without significantly increasing trainable parameters.
• For instance, ShuffleNetV2 with GNTF achieved a 3.3% accuracy improvement on CIFAR-10-C with fewer parameters.

Conclusions:

• The proposed GNTF method effectively enhances the performance of lightweight CNNs on degraded data.
• This approach offers a viable solution for deploying reliable vision services on IoT devices in complex outdoor environments.
• GNTF contributes to more robust and efficient AI on edge devices.