Formulates quadratic ReLU replacement as a linear separation problem in lifted space, with exact conditions for calibration-lossless replacement and convex relaxations for approximate cases, achieving plaintext accuracy at lower cost under CKKS.
Optimized layerwise approximation for efficient private inference on fully homomorphic encryption
2 Pith papers cite this work. Polarity classification is still indexing.
2
Pith papers citing it
verdicts
UNVERDICTED 2representative citing papers
Power-Softmax is a new HE-compatible attention variant that permits training and inference of billion-parameter polynomial LLMs with performance matching standard transformers.
citing papers explorer
-
Power-Softmax: Towards Secure LLM Inference over Encrypted Data
Power-Softmax is a new HE-compatible attention variant that permits training and inference of billion-parameter polynomial LLMs with performance matching standard transformers.