Data Fault Injection for MLaaS (Machine Learning as a Service).
This framework allows injecting data faults into the input data provided to Machine Learning cloud services and then assessing the robustness of these services when exposed to those faults based on a set of available metrics.
The framework implements interfaces for using machine learning APIs from three major cloud providers. Currently, only computer vision services are implemented:
- Amazon Rekognition (AWS): label detection, text detection, violence detection, nudity detection, celebrity recongnition, face detection
- Vision AI (Google Cloud): label detection, text detection, violence detection, nudity detection, face detection
- Vision Service (Microsoft Azure): label detection, nudity detection, face detection
Data Faults are problems in the input data of a system that may arise from external data collection (e.g., sensors, cameras) or data manipulation routines [1]. The framework implements a total of 11 data faults that can be injected into images.
The available image faults in the framework and their parameters are listed below. Their implementations rely on open-source Python packages, namely imagecorruptions, numpy, pillow, and scikit-image.
- Brightness Change
- Chromatic Aberration
- Gaussian Blur
- Motion Blur
- Zoom Blur
- Gaussian Noise
- Grayscale
- Missing Pixels
- Salt & Pepper Noise
- Contrast
- Pixelation
- Condensation
- Fog
- Frost
- Rain/Snow
mlaas-fi requires a dataset to run experiments. Valid image datasets are are tarball files that should expand to a single directory with the dataset images.
Five datasets were used in our study to evaluate the robustness of MLaaS services to data faults:
- Celebrity Recognition: A random sample of 500 aligned and cropped images from the CelebA dataset [2].
- Label Detection: A random sample of 500 images from the 2017 validation set of the COCO object detection dataset [3].
- Nudity Detection: A random sample of 500 images from the test set of the NudeNet dataset [4].
- Violence Detection: random A sample of 198 images of child soldiers and police violence from the Human Rights UNderstanding (HRUN) dataset [5]. All child soldier images have the presence of weapons.
- Text Detection: A random sample of 398 images captured in different environments from the KAIST Scene Text Database [6]. Only images with English text only were selected.
[1] Nurminen, Jukka K., et al. "Software framework for data fault injection to test machine learning systems." 2019 IEEE International Symposium on Software Reliability Engineering Workshops (ISSREW). IEEE, 2019.
[2] Liu, Ziwei, et al. "Deep learning face attributes in the wild." Proceedings of the IEEE international conference on computer vision. 2015.
[3] Lin, Tsung-Yi, et al. "Microsoft coco: Common objects in context." European conference on computer vision. Springer, Cham, 2014.
[4] “NudeNet Classifier Dataset v1 : Https://Github.com/bedapudi6788/ : Free Download, Borrow, and Streaming.” Internet Archive, archive.org/details/NudeNet_classifier_dataset_v1. Accessed 2 May 2021.
[5] Kalliatakis, Grigorios, et al. "Detection of human rights violations in images: Can convolutional neural networks help?." arXiv preprint arXiv:1703.04103 (2017).
[6] SeongHun Lee, Min Su Cho, Kyomin Jung, and Jin Hyung Kim, "Scene Text Extraction with Edge Constraint and Text Collinearity Link," 20th International Conference on Pattern Recognition (ICPR), August 2010, Istanbul, Turkey.