I’m presenting our paper "Trojan Source: Invisible Vulnerabilities" coauthored with @rossjanderson at USENIX Security at 1:30pm PT today. Come check it out! https://t.co/ITubxfutJa
Uncommon Unicode encodings can be used to manipulate the results of search engines and their chatbot counterparts. If an adversary can manipulate a user into copying a search term into a search engine, they can leverage this behavior to control the results.
Imperceptible adversarial examples have now arrived for NLP. Unicode can be used to perturb text encodings in a way that targets ML models without affecting humans. Our techniques can be used to attack systems ranging from toxic content detection to machine translation.
I’m presenting this work in a paper titled "Bad Characters: Imperceptible NLP Attacks" coauthored with @iliaishacked, @rossjanderson, and @NicolasPapernot at IEEE S&P at 10:40am PT today. Come check it out.