Nathaniel Fruchter

@nhfruchter

Privacy engineering at Google ( '18, '16 before). All opinions my own.

The weird part of the Internet
Vrijeme pridruživanja: listopad 2007.

Tweetovi

Blokirali ste korisnika/cu @nhfruchter

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

  1. 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
  2. 29. sij

    And that's a wrap! passing the torch to and as next year's chairs for 2021 (Oakland). Thanks everyone!

    Poništi
  3. 29. sij

    Remaining problems: - Units! How do we standardize / compare measurement of abuse. It's an open question for the audience. - Systems! How can we combine different approaches into a unified system? - Reporting! Trustworthy signals from user-sourced reports.

    Prikaži ovu nit
    Poništi
  4. 29. sij

    Philosophy of measurement: Goodhart's Law: "When a measure becomes a target, it ceases to be a good measure." Uncertainty principle: "Any abuse signal can be used for measurement or enforcement, but not both."

    Prikaži ovu nit
    Poništi
  5. 29. sij

    The various approaches.

    Prikaži ovu nit
    Poništi
  6. 29. sij

    The gold standard: human labeling. FB uses stratified sampling to oversample suspicious activity due to base rates of abuse content.

    Prikaži ovu nit
    Poništi
  7. 29. sij

    Approaches to measurement: 1. Count what you blocked. Are spikes good or bad? 2. Have users do the labeling (reporting). Reporters have inherent biases though (not necessarily bad), and data can be a bit unreliable.

    Prikaži ovu nit
    Poništi
  8. 29. sij

    Benefits of measurement: track progress/regressions, data for models, deep dives, retroactive takedown.

    Prikaži ovu nit
    Poništi
  9. 29. sij

    Measuring abuse: the dark ages. - Prioritization: who yells the loudest? - Impact: blocked actions over time. Next age ("the renaissance"): - Prioritization: volume x relative harm = rank - Impact: how much did you *not* block?

    Prikaži ovu nit
    Poništi
  10. 29. sij

    What's the relative impact of 1 spam post to 1 account takeover? We don't necessarily know.

    Prikaži ovu nit
    Poništi
  11. 29. sij

    Adult content, merchandise spam, scams... there's a lot going on out there. But it's hard to measure. Issues: no common units, unknown unknowns, adversarial response, category imbalance, no good concept of ground truth.

    Prikaži ovu nit
    Poništi
  12. 29. sij

    Speaking of metrics, though: how do we figure out what to do? How do we measure the impact of our work in anti-abuse? How do we prioritize?

    Prikaži ovu nit
    Poništi
  13. 29. sij

    "Sometimes I wished I worked on ad ranking..." It's easier to determine what works in ads. [Or is it? Today's earlier talk about metrics in ads...]

    Prikaži ovu nit
    Poništi
  14. 29. sij

    Now: David Freeman from on the "abuse uncertainty principle" and measuring online abuse.

    Prikaži ovu nit
    Poništi
  15. 29. sij

    Incidents are opportunities to improve! This relatively small program created relatively massive dividends.

    Prikaži ovu nit
    Poništi
  16. 29. sij

    Reservists gave back 76 days worth of time to the team, which allowed for breathing room and more proactive work to be done.

    Prikaži ovu nit
    Poništi
  17. 29. sij

    Reservist model = pool of auxiliary crisis managers (voluntary!) that has prior experience. Incentives for reservists: engagement, training, collaboration, gained knowledge.

    Prikaži ovu nit
    Poništi
  18. 29. sij

    What if we analyzed the incident data for trends? Netflix created forecasts of future incident load, as there are a lot of (seeming) correlates about time, quarter, area of incident, etc.

    Prikaži ovu nit
    Poništi
  19. 29. sij

    - Raised the bar on what an incident was - built a high-level, distributed process for response. Turned team into an escalation point. - Efficiencies: crisis mgmt automation tools for support/admin. This reduces cognitive load on responder and time to assemble.

    Prikaži ovu nit
    Poništi
  20. 29. sij

    Obviously don't want this to happen again, but scaling incident response is hard! Goals: detect quickly, respond effectively, and limit the blast radius from the incident.

    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:

    ·