Accurate identification of misuse is a key factor in determining appropriate ways to protect systems. Modern intrusion detection systems often use alerts from different sources such as hosts and sub-networks to determine whether and how to respond to an attack. However, alerts from different locations should not be treated equally. We propose improving and assessing alert accuracy by incorporating an algorithm based on the exponentially weighted Dempster-Shafer (D-S) Theory of Evidence. Our approach uses D-S theory to combine beliefs in certain hypotheses under conditions of uncertainty and ignorance, and allows quantitative measurement of the belief and plausibility in our detection results. Our initial evaluations on the DARPA IDS evaluation data set show that our alert fusion algorithm can improve alert quality over those from Hidden Colored Petri-Net (HCPN) based alert correlation components installed at the demilitarized zone (DMZ) and inside network sites. Due to alert confidence fusion in our example, the detection rate rises from 75% to 93.8%, without adversely affecting the false positive rate.