Privacy implications of IFT: adversarial fonts

[svgeesus]

From Robin on Mastodon

gotcha, I'm glad that's already on your radar. I still have a bunch of questions about the implications of all of this.

1. Where is it shown that an adversarial font would have prohibitively high latency?
2. Doesn't the data show that the main way to mitigate this risk is randomness (I'm not counting font group size because that is up to the server/producer) and yet the spec doesn't seem to require that?
3. It's still unclear to me how ligatures interact with this and if they let you capture bigram data and how that would affect the simulation results?
4. What was the methodology for collecting the IFT performance analysis data set? Is it suitable and representative of what's needed for privacy analysis?
5. Was the use of characters outside the main language (e.g. "emojis") considered as part of the fingerprinting?