Lea Kissner

@LeaKissner

Chief Privacy Officer , formerly Global Lead of Privacy Technology for . Privacy eng, crypto, security, and building for respect. she/her

California, USA
Vrijeme pridruživanja: travanj 2017.

Tweetovi

Blokirali ste korisnika/cu @LeaKissner

Jeste li sigurni da želite vidjeti te tweetove? Time nećete deblokirati korisnika/cu @LeaKissner

  1. Prikvačeni tweet
    16. sij

    2020 is here and so is PEPR 2020! We had a *lot* of requests to move to a standalone event that wasn't in August. So this year: May 11-12, in Santa Clara. The CFP is up, so send us your talk proposals about privacy engineering, practice, and respect.

    Prikaži ovu nit
    Poništi
  2. proslijedio/la je Tweet
    30. sij

    This is your regular reminder to people in the tech industry that the choices you make in your job matter.

    Poništi
  3. 30. sij

    My answer to "what's the worst that can happen?" is "I have an *extremely* good imagination"

    Prikaži ovu nit
    Poništi
  4. 30. sij

    This is why my coworkers have learned not to ask me "what's the worst that can happen?"

    Prikaži ovu nit
    Poništi
  5. 30. sij

    That's the end of . Already looking forward to , which I'll be co-chairing with . So get your thinking caps on, send PEPR talks to PEPR (open now!) and start thinking about awesome Enigma talks. I can't wait to see them.

    Poništi
  6. 29. sij

    Q: did you do any analysis on which labeling approaches work better across languages/cultures? A: I don't know offhand, I might be able to follow up

    Prikaži ovu nit
    Poništi
  7. 29. sij

    Q: can you speak some more about how you balance the measurement/enforcement tradeoff? do you ever hold off on blocking to get better measurement? A: with a new problem measure-first. for more mature problems it's either working in parallel or series on measurement/enforcement

    Prikaži ovu nit
    Poništi
  8. 29. sij

    Q (again): ToS isn't very clear A: I'm not a lawyer

    Prikaži ovu nit
    Poništi
  9. 29. sij

    Q continued: Also, it seems like if a company is more transparent about the formula it would help users but would also educate the adversary. A: try to make the standards very public and in clear language We've been publishing some of these measurements for years.

    Prikaži ovu nit
    Poništi
  10. 29. sij

    Q: It seems like you would have a lot of data about abuse and ways to measure abuse. Someone like me would assume you'd make some guidelines and share them across multiple platforms.

    Prikaži ovu nit
    Poništi
  11. 29. sij

    A: high-level strategy is to use unsupervised learning for training to train supervised models for detection, so adversaries don't get immediate feedback. Continuous learning/retraining, up to on-line trained models that refresh hourly.

    Prikaži ovu nit
    Poništi
  12. 29. sij

    Q: You mentioned that you use ML to label some of this bad behaviour. Can you touch on some of the strategies you use, given that adversaries will be... adversarial and work around it?

    Prikaži ovu nit
    Poništi
  13. 29. sij

    Q: when you are eliciting expert judgement how do you account for different opinions? A: the experts are people within the company and they work according to the policy made by policy team which has to take in different perspectives

    Prikaži ovu nit
    Poništi
  14. 29. sij

    Q : when you're oversampling bad stuff to feed to raters, how do you know that you're not missing false negatives. A: there's a talk at <at>scale to point you at, but the upshot is that as long as your sample function is better than random you still get improvement

    Prikaži ovu nit
    Poništi
  15. 29. sij

    A: There's work on automated labeling (if a computer can catch it, don't show to human), break stuff up into different parts e.g. differnet people looking at different parts of an account

    Prikaži ovu nit
    Poništi
  16. 29. sij

    Q: you mentioned problems about scaling the human labelers, but you didn't mention the impact on people of looking at horrific things. Do you think there is any hope that this can trend downwards?

    Prikaži ovu nit
    Poništi
  17. 29. sij

    Also validate your metrics e.g. by looking on the black market: how much does a like/comment cost? Challenge: universal adapter so I can compare different kinds of abuse and false positives/false negatives so can effectively choose what to work on.

    Prikaži ovu nit
    Poništi
  18. 29. sij

    Drive down one while measuring with the next metric. [ That is very clever ]

    Prikaži ovu nit
    Poništi
  19. 29. sij

    Sorry I really can't type out these slides

    Prikaži ovu nit
    Poništi
  20. 29. sij
    Prikaži ovu nit
    Poništi
  21. 29. sij
    Prikaži ovu nit
    Poništi

Čini se da učitavanje traje već neko vrijeme.

Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.

    Možda bi vam se svidjelo i ovo:

    ·