This is less a hint about improving accuracy than one about not deteriorating the accuracy of your junk mail filter.
Most email spam filters classify spam based on word frequency. When you train the filter, you give the filter a list of bad words. If a particular bad word comes up frequently it increases the likelihood that that email is spam. Then when an email comes in with enough bad words it is classified as spam.
The .gif emails that are currently going around do not get filtered well by spam filters. These emails contain two parts: the first part, the .gif, contains the text that the filter would normally learn to trigger off. Since this text is in the image, the filter can't see it. Instead the filter sees the second part of the email: a list of random phrases and words. The filter picks up these words and calculates a spam score from them.
The best thing to do with these spams is simply delete them and not try to train your filter with them. If you do have your filter learn the words in these messages, it will only be learning common words, which will skew its results. Use a rule to highlight these emails or to move them to your spam folder, as explained in this hint; it will maintain the integrity of your spam filter.
[kirkmc adds: For what it's worth, I had not paid attention to the attachment spams that still reached my inbox recently, but while editing this hint, looked at a few and see that they contain png images now. (The gifs are all filtered.) So if you do use a rule, such as in the recent hint linked above, think about creating another one for png format graphics. It's only a matter of time before spammers start using jpgs and other formats, though...]
Mac OS X Hints
http://hints.macworld.com/article.php?story=20061031221923601