Continual malware detection requires adaptation without compromising previously learned detection capabilities. However, preserving prediction consistency can limit the correction of earlier errors. Our diagnostics show that logit matching is associated with reduced predecessor-error correction in the evaluated Positive Congruent Training (PCT) configuration, while correctly detected malware nearest the raw decision boundary has the highest mean next-update flip rate. We introduce Boundary-Guided Residual Correction (BGRC), which freezes each updated PCT backbone and fits a nonnegative residual of compactly supported atoms and a global intercept. Residual fitting combines current-data supervision, coefficient regularization, and selective margin preservation, supported by a bounded memory prioritizing correctly detected malware near the raw boundary.
We evaluate BGRC on ELSA and TESSERACT extended against PCT, replay, multilayer perceptrons (MLPs) trained only with cross-entropy (CE), and correction-time-matched PCT--CE. At the primary calibration target, BGRC improves next-quarter F1 over PCT by up to 5.400 percentage points (pp). On ELSA, it achieves the lowest mean backward malware negative-flip rate (NFR) among the evaluated methods while nearly matching sequential MLP fine-tuning in mean F1. On TESSERACT extended, it reduces backward NFR by 3.061 pp relative to PCT--CE at a cost of 2.702 points in F1.
These results demonstrate that BGRC can improve the detection--retention trade-off in continual malware detection.