Frequency-Aware Continual Learning Boosts Smart Contract Security

Smart contract security is a critical concern, and detecting vulnerabilities with Large Language Models (LLMs) faces several interconnected challenges. As new vulnerability types emerge, LLMs need to adapt without costly full retraining, which can lead to forgetting previously learned vulnerabilities. Furthermore, managing multiple specialized adapters for different tasks and consolidating them into a single, efficient model for real-time inference is complex. Researchers have developed an integrated three-stage framework to tackle these issues. The first stage, Frequency-Aware Low-Rank Adaptation (FA-LoRA), enables parameter-efficient adaptation by operating in the Fourier domain. The second stage, Forget-Aware Replay (FAR), uses frequency gates to identify and prioritize vulnerable knowledge for rehearsal, preventing catastrophic forgetting. Finally, Anchor-Protected Progressive Merging (APPM) consolidates multiple adapters into a single, robust model, ensuring strong generalization and efficient deployment. This framework has been experimentally validated on blockchain ecosystems, demonstrating its effectiveness in maintaining high detection accuracy while managing the complexities of continuous learning and model deployment for smart contract security.