Stealing Split Learning Bottom Models by Recovering Embedding Geometry
VENOM first learns a contrastive space over server-observed embeddings, then builds a neighborhood graph and trains a surrogate bottom model to match targets and respect local geometry via a neighbor-matching loss alongside pointwise and feature-shape alignment to preserve the relational structure that defenses fail to...
Qinbo Zhang, Yanhang Shi, Ziying Zhang et al.
· 1 citation