“Theoretical research has demonstrated that it is possible to introduce hidden, undetectable backdoors into machine learning models.”