Gradients Increase Text Leakage in Split Learning by 3.17%
October 1, 2026
In split learning, an attacker can rebuild 94.20% of client tokens from activations alone, increasing to 97.38% when gradients are included. For 32-token documents, gradient access increases exact document reconstruction from 13.71% to 37.77%.
HOW THIS AFFECTS YOU
●
builderYou must assume that gradient exchange in split learning significantly increases the risk of data leakage.
●
policyThis highlights a critical privacy vulnerability in distributed training architectures.