image processing, neural network inference, adversarial machine learning

This research thoroughly analyses the effects of various image distortions on different neural network architectures, utilizing the ImageNet validation dataset. The study reveals that distortions such as blur, noise, JPEG compression, and grid patterns significantly elevate error rates. Contrast change yielded inconsistent impacts, only occasionally increasing error rates, while rescaling the image drastically (more than five times) led to substantial model performance degradation, although the images became difficult for even humans to recognize at such scales. Among the architectures tested, EfficientNet models demonstrated superior robustness, which can be attributed to their scalable input sizes. The Vision Transformer (ViT), trained on the extensive JFT-300M dataset, also showcased notable resilience to distortions. Other architectures like ConvNeXt, Swin, and ResNetD were more susceptible to the tested distortions. MobileNetV3-based models seem to provide an exceptionally good ratio between the robustness and model size, but this could also be caused by the fact that models from this family have inherently higher base error rates, which intrinsically allows for a larger margin of error.

This research thoroughly analyses the effects of various image distortions on different neural network architectures, utilizing the ImageNet validation dataset. The study reveals that distortions such as blur, noise, JPEG compression, and grid patterns significantly elevate error rates. Contrast change yielded inconsistent impacts, only occasionally increasing error rates, while rescaling the image drastically (more than five times) led to substantial model performance degradation, although the images became difficult for even humans to recognize at such scales. Among the architectures tested, EfficientNet models demonstrated superior robustness, which can be attributed to their scalable input sizes. The Vision Transformer (ViT), trained on the extensive JFT-300M dataset, also showcased notable resilience to distortions. Other architectures like ConvNeXt, Swin, and ResNetD were more susceptible to the tested distortions. MobileNetV3-based models seem to provide an exceptionally good ratio between the robustness and model size, but this could also be caused by the fact that models from this family have inherently higher base error rates, which intrinsically allows for a larger margin of error.