Think Security, Write Security

Scholarly Communication

Listen to this interview of Nicholas Boucher, PhD, Department of Computer Science and Technology, Cambridge University, UK. We talk about his coauthored paper Bad Characters: Imperceptible NLP Attacks (SP 2022) — and check out, too, Nicholas's presentation of the paper here.

Nicholas Boucher: "Maybe what is interesting about the security domain is that, oftentimes, in these attack papers, you start with a hypothesis, but it's an hypothesis already informed by some result you've observed in the wild — so, you've seen some sort of system — or, to be concrete, in our case, we saw people switching between alphabets on keyboards, and that enabled us to notice how such an action could interact with the language models quickly growing in popularity — and it is at that point that a security researcher will say, 'Wow, I have something here. I know that this is a vulnerability.' But then the questioning begins, like, how to frame the vulnerability, that is, how to turn one specific example (which the researcher has a strong feeling really is a vulnerability) and uplevel it to something larger. Because that is when, in my opinion, the researcher's starting to ask very fruitful questions."

Learn more about your ad choices. Visit megaphone.fm/adchoices