Spam Reporting @Fermilab

What are we doing?

• We have been using SpamAssassin to tag messages with a "spam score" that mail programs can use to filter messages.
• We recently activated the Bayesian Filtering module of SpamAssassin.

What is Bayesian Filtering?

• Unlike the rest of the SpamAssassin local rules, as opposed to the network based rules, Bayesian Filtering is an adaptable spam filtering method.
• It does not use specific predefined tests.
• It relies on a training program to analyze known spam emails and known non-spam emails and generates databases with counts of the frequencies of each word appearing in spam and non-spam emails.
• When scanning a message, it combines the scores for all the words in the message and comes up with a percentage that indicates the likelihood a message is spam. This is then used to add or subtract points to the normal SpamAssassin score.

What messages are we training with?

• For most of the training, we are using auto-learning.
• This takes messages with a normal SpamAssassin score < 0.1 or greater than 12 and trains the Bayesian Filter databases on those messages.
• We also take user submissions via the Spam Reporting Form and train on them.
• The submitted spam is minimally reviewed by a person just to make sure anything that is obviously not spam is not accidentally submitted.

Links:

• Spam Reporting Page.
• Sample spam.
• Sample non-spam.