Synapse
⌘+K
Synapse
PulseExploreClubsResearchersJournals
Instagram
HomeClubsExplore
April 23, 2026Open Access

Cross-Domain Vulnerability Detection using LLMs: Knowledge Transfer from Contemporary Software to AI/ML-Specific Software

View Full Paper
Ask AI
Bookmark
Share

Authors

MSMiltiadis SiavvasIKIlias KalouptsoglouDKDionysios Kehagias

Discussion

Loading...

Member takes

Overview

Cross-domain vulnerability detection reveals the importance of domain adaptation in AI/ML software security.

Key Points

  • Examine if vulnerability detection models trained on contemporary software can accurately detect vulnerabilities in AI/ML-specific software.
  • Construct a dataset of vulnerable and non-vulnerable Python functions from contemporary and AI/ML-specific subsets.
  • Fine-tune CodeBERT and CodeGPT on the contemporary subset and evaluate on the AI/ML subset.
  • Perform gradual adaptation by adding varying percentages of target domain samples and analyze representation drift with UMAP.
  • AI-based vulnerability detection models trained on contemporary software performed poorly on AI/ML-specific software without domain adaptation.
  • A moderate amount of domain-specific data (as low as 5%) improved detection accuracy significantly.
  • Unique characteristics of AI/ML-specific software necessitate tailored approaches for effective vulnerability detection.

Cite This Study

Siavvas et al. (2026) studied this question.

synapsesocial.com/papers/69e9bb6285696592c86ed193https://doi.org/10.5281/zenodo.19679663
View Full Paper
Ask AI
Bookmark
Share