MS student, Tianjin University
1 paper at NeurIPS 2025
This paper proposes LoSplit, a training-time defense that detects and mitigates graph backdoors by dynamically analyzing early loss divergence.
