The present invention discloses a network-based method and system for filtering a junk message. The method includes the following procedures: word-segmenting a target message to obtain characteristic elements within the target message; obtaining probabilities of the characteristic elements in the target message according to a predefined keyword library, the keyword library being updated periodically; and accumulating the probabilities of the characteristic elements in the target message, comparing accumulated probabilities with a predetermined threshold, and concluding the target message is a junk message if the result is greater than the predetermined threshold. Using a predefined keyword library for verification of a target message avoids time delay and system resource waste caused by massive and complicated computations during message verification in existing technologies, thus satisfying the requirement of real-time filtering of junk messages.