How Corrections Are Used

A correction is what happens when someone tells SpamFoo it got a message wrong. This page explains what SpamFoo does with a correction on your server and what it sends to SpamFoo's servers.

Kinds of Corrections

  • Spam corrections: Marking a message as spam or phishing, or rescuing a message from spam. Phishing can only be reported from the SpamFoo dashboard. See Correct a Wrong Result.
  • Classification corrections: Moving a message to a different Email Classification, such as from Primary to Updates. See Correct a Wrong Result.

How Corrections Reach SpamFoo

  • SmarterMail: When users mark spam, rescue mail, or move messages between classifications, SmarterMail sends the correction to SpamFoo for them. See SmarterMail Overview.
  • Other mail servers and custom software: The software must send each correction to SpamFoo Client. See Send Corrections Back.
  • The SpamFoo dashboard: You can report a message from its details. In the dashboard, a correction is called a report.

What Happens on Your Server

  • SpamFoo adds the correction to that user's pattern memory, and it takes effect immediately.
  • The correction only affects the user who made it.
  • This works even when SpamFoo's online services can't be reached.

See How Personalization Works for how pattern memory decides future mail.

What Is Sent to SpamFoo

Corrections also help train the models that ship to every SpamFoo installation. SpamFoo only reports a correction when its own models got the message wrong. It doesn't report one when:

  • The user's own past corrections already decided the message.
  • The message was already classified the way the user chose.

A Phishing report from the SpamFoo dashboard is sent even when SpamFoo's models already marked the message as spam, because it tells SpamFoo something new.

Sent by Default

By default, a reported correction contains mathematical data about the message, not the message itself:

  • A fingerprint of the message: a list of numbers that captures what it is about. The words can't be read back from it.
  • Signals SpamFoo measured, such as sender check results, link counts, and whether there were attachments.
  • What SpamFoo decided, what the user chose instead, whether the user reported it as phishing, and how sure the models were.
  • The sender's domain and the IP address of the server that sent the message.
  • A hashed form of the mailbox address that made the correction. Hashing turns the address into a fixed code instead of readable text.

The subject, the message text, and the sender's full address aren't included.

Sharing Message Content

An administrator or the owner of the mailbox can choose to also share a sanitized copy of the message text with corrections. This helps train the models better than the mathematical data alone.

  • Off by default. Nothing beyond the mathematical data is sent until someone turns it on.
  • Corrections from SmarterMail or other software: The mail server decides, for each correction, whether the user agreed to share content. If it doesn't say, SpamFoo treats the answer as no.
  • Reports from the SpamFoo dashboard: A system administrator controls this for the whole server with Share Correction Content in the Message Content card of System > Settings. It only works while Store Message Content is on. The first time a system administrator reports a message, SpamFoo asks whether to share.
  • Sanitizing: Before sending, SpamFoo replaces email addresses, phone numbers, IP addresses, Social Security numbers, credit card numbers, and common names with placeholders, and scrambles other numbers. Unusual details, such as a name SpamFoo doesn't recognize, may still get through.
Warning Do not turn on content sharing for users who have not agreed to it.

Automatic Training Reports

An administrator can also turn on automatic training reports when starting SpamFoo Client. SmarterMail turns them on when its own setting for reporting to SpamFoo is on.

  • SpamFoo then sends the same kind of mathematical data for some messages it was unsure about, and for a small random sample of ordinary mail.
  • These reports never include message text.
  • They are off unless turned on.

When SpamFoo's Servers Can't Be Reached

  • The correction still takes effect on your server right away.
  • The report waits in a queue on your server and is retried for a while. If it still can't be delivered, it is dropped.
  • Mail delivery isn't affected.

What to Expect

  • Spam Rules are exact. A rule for a sender always applies, unless a stronger rule from an administrator matches. See Manage Your Rules.
  • Pattern memory works by similarity. One correction is often enough for the same sender. For mail from many different senders, several corrections of similar messages work better than one.
  • Model updates improve results for every user, including users who never send corrections. They arrive with regular model updates. See Update and Back Up.