Alpha-Split Method Fixes Budget Collapse in Federated Speech-LLMs
September 11, 2026
Standard per-layer differential privacy clipping fails in speech-LLMs due to norm imbalances between acoustic encoders and language decoders. The proposed alpha-split method uses two independent parameter pools to normalize encoder and LLM updates, preventing cross-component budget collapse and improving word error rates.
HOW THIS AFFECTS YOU
●
builderThis provides a path to training privacy-preserving speech models without catastrophic loss in utility.
●
researcherYou should use dual-pool allocation when applying DP to architectures with highly disparate component norms.