The release of the spam filtering system Rspamd 3.9 has taken place, providing tools for assessing messages based on various criteria, including rules, statistical methods, and blacklists, used to determine the final weight of a message for making a blocking decision. Rspamd supports almost all capabilities implemented in SpamAssassin and has features that allow email filtering to be about 10 times faster than SpamAssassin while providing better filtering quality. The system's code is written in C and is distributed under the Apache 2.0 license.
Rspamd is built using an event-driven architecture and is initially designed for use in high-load systems, allowing the processing of hundreds of messages per second. The rules for identifying spam characteristics are highly flexible and can be expressed in simple forms using regular expressions, while more complex situations can be written in Lua. Functionality expansion and the addition of new types of checks are implemented through modules that can be created in C and Lua. For example, there are modules available for sender verification using SPF and sender authentication via DKIM, as well as querying DNSBL lists. domain An administrative web interface is provided to simplify configuration, rule creation, and statistics tracking.
In the new version:
- Improvements have been made to the Bayesian classifier settings. The default window size has been reduced from 5 to 2 words, resulting in enhanced performance and a 4-fold reduction in storage space consumption without degrading spam classification levels. A utility called "rspamadm classifier_test" has been provided to test the classifier's operation with different settings.
- A GPT module has been added, utilizing the OpenAI GPT API for text classification via requests to large language models such as GPT-3.5 Turbo and GPT-4. The spam classification accuracy with the new module is lower than that of the Bayesian classifier, but its advantage lies in the fact that it does not require pre-training and can take context into account in messages, while effective operation of the Bayesian classifier requires high-quality and balanced training of the engine. Besides directly identifying spam in messages, the GPT module can also be used to train the Bayesian classifier.
- The ability to share the known_senders and replies modules has been implemented to mark verified senders based on the criterion that replies have been sent to them before.
- Dynamic adjustment of message sending rate limits related to a single sender or recipient is disabled by default. IP address.
Source: opennet.ru
