Bingdao Feng

This paper proposes LoSplit, a training-time defense that detects and mitigates graph backdoors by dynamically analyzing early loss divergence.