Security bug reports can describe security critical vulnerabilities in software. Bug tracking systems may contain thousands of bug reports, where relatively few of them are security related. Therefore finding unlabelled security bugs among them can be challenging. To help security engineers identify these reports quickly and accurately, text-based prediction models have been proposed. These can often mislabel security bug reports due to a number of reasons such as class imbalance, where the ratio of non-security to security bug reports is very high. More critically, we have observed that the presence of security related keywords in both security and non-security bug reports can lead to the mislabelling of security bug reports. This paper proposes FARSEC, a framework for filtering and ranking bug reports for reducing the presence of security related keywords. Before building prediction models, our framework identifies and removes non-security bug reports with security related keywords. We demonstrate that FARSEC improves the performance of text-based prediction models for security bug reports in 90% of cases. Specifically, we evaluate it with 45,940 bug reports from Chromium and four Apache projects. With our framework, we mitigate the class imbalance issue and reduce the number mislabelled security bug reports by 38%.