Retrofit: Continual Learning with Controlled Forgetting for Binary Security Detection and Analysis
This episode discusses the 'RETROFIT' paper, a continual learning method designed for security analysis like malware detection and binary summarization. It solves the problem of catastrophic forgetting—where models lose knowledge of old threats when learning new ones. RETROFIT achieves this without storing historical data, using low-rank updates and confidence-guided arbitration to maintain performance while achieving substantial gains over existing methods.