This is interesting! Can’t you just filter this range of unicode characters out of your apps input before it hits the model? Most models aren’t exposed to users as direct inference, it’s wrapped in an application. I guess there’s some implication for data poisoning, but realistically datasets should be fed through similar filtering to look for adversarial inputs.
@joseph_thacker
8 ай бұрын
Yes, but that's like saying "all apps should sanitize user input against xss and sanitize sql queries against injection". I've been doing bug bounty hacking for more than 4 years with over 1000 reports. Most companies still struggle to do those 2 things well. This issue feels like a simialr problem. Every app developer has to fix it. I'd rather scrub the training data of any unicode tags so future models don't even understand them. There's very little use.
@endlessvoid7952
8 ай бұрын
@@joseph_thackervery true, just because it should be filtered doesn’t mean everyone will do it or even be aware that they need to 😢
Пікірлер: 3