AI supply chain risks
Much like software development, AI increasingly leverages pre-trained models and crowd-sourced data to expedite development. ML frameworks such as Keras provide straightforward access to a range of pre-trained models and sample datasets. Additionally, models and datasets are readily available on various platforms, including GitHub repositories, specialized websites, and dedicated communities such as Kaggle. TensorFlow Hub, Model Zoo, and PyTorch Hub are model repositories containing hundreds of pre-trained models ready to be used directly or via transfer learning. Hugging Face is becoming a popular marketplace that simplifies the acquisition of state-of-the-art models.
However, this convenience comes with its own set of risks. Utilizing pre-trained models and datasets from external sources exposes organizations to potential vulnerabilities.
These risks range from malware, model poisoning, and tampering to data poisoning, model biases, intellectual property...
