EvoSec 2025W Lecture 14

From Soma-notes

Readings

Discussion Questions

  • How quickly can you verify that another person knows the same story that you do? Can someone who doesn't know the story fool you easily?
  • Are there stories that are known to friends and family but that are not written down anywhere?
  • Could any social media or financial site "tell a story" about your interaction with them that isn't generally known?

Notes

Lecture 14
----------

If you haven't finished your early lit review yet, please get it to me by the end of the week
 - for those who have submitted, I will get you feed back soon!

G1
 - protects against non-targeted attacks but may be more vulnerable to targeted attacks, with LLMs
 - if the story's generic may be easy to guess, needs to be different
 - sibling stories!
 - your social media feed tells a lot about you, as does your record of financial transactions, can even know you're pregnant before you do!
 - narrative authentication allows for better two-way trust, potentially
 - "sharing a story" could actually involve lots of differences in how the story went
 
G2
 - making up details (and waiting for the "wait a minute" remark) can help you figure out whether someone shares the same story
 - oral traditions in aboriginal societies
 - "mandela effect" - misremembering stories collectively
 - meaningful interactions with a site can lead to interesting stories

G3
 - LLM internal model would be probability based, could be hacked
   - social engineer details out of people
   - easiest attacks target people
 - new security issue, hacking LLMs to bypass authentication!
 - more complex narratives are more secure but harder to remember
 - if it's too much work to create authentication credentials users won't do it

G4
 - could LLMs guess based on context clues?
   - 12 year old male with an interest in baseball -> probably a predictable birthday cake
   - but made-up details could help you catch impersonators but could fool legit people
   - would need details that people wouldn't forget
 - instagram knows who you are creeping
 - link history can tell a lot, e.g., tell e-commerce site what you are interested in


How do we recognize and trust each other?
 - and what would it mean for the same approach to apply to computational systems

good narratives implicitly are based on a model of their audience
 - biased towards "interesting" bits with an "interesting" structure
 - interesting => high information content, not knowable just from priors