Manipulating Machine Learning Systems by Manipulating Training Data

Recent work has identified that classification models implemented as neural networks are vulnerable to data-poisoning and Trojan attacks at training time.