LabelCaptcha can produce impossible captchas when irrelevant words are maliciously added to the known words list about lc-core HOT 1 OPEN

Thank you for the analysis!

I can think of a way to mitigate this, but there may be better options for solving this: Record the number of succeeded and failed attempts of each known word per image. Known words with a high number and proportion of failed attempts should be removed from the known words list

Yup; this is a good solution for the problem. However, this should be only applied to words/images that have been promoted to known by the CAPTCHA system, so that definitely known images don't get demoted by wrong answers.

Proposal: Instead of the two category system, we can migrate to a four category system:

• known : these images are invariants, never demoted
• unknown : these are images that need to be solved by the system
• solved : these images were promoted from unknown by the system when they achieved good success rate. They may be demoted to held if their success rate drops.
• held: these images are held for manual review, and won't be utilised by the system

Further, we should mixup the order of known and unknown words randomly. And, sometimes, use known words only.

from lc-core.