Machine-learning models vulnerable to undetectable backdoors: new claim

For some rather unexciting definitions of “backdoor” anyway.

If an email service notices big amounts of spam evading their ML classifier, they’ll have to change their systems just like they would have done pre-ML.

If a bank notices a big amount of bad loan applications being approved by their ML-classifier, they’ll have to change their systems just like they would have done pre-ML.

Nobody is going to be popping shells with this. The “worst case” scenarios are already happening on a regular basis with traditional systems without any backdoors at all.