ATLAS: Automated Polynomial Approximation for FHE Transformer Inference
A new framework called ATLAS automates the configuration of polynomial approximations for transformer models under fully homomorphic encryption (FHE), reducing setup time from days to one hour. FHE enables private inference but requires replacing non-linear operations like softmax and normalization with polynomial approximations compatible with the CKKS scheme. The multiplicative depth of these approximations dominates inference cost. Existing frameworks use manually configured hyperparameters applied uniformly across all layers, which is suboptimal because different layers tolerate different error levels. ATLAS allows per-layer configuration to exploit this variability, reducing latency without sacrificing accuracy. The system automatically searches for optimal approximation hyperparameters, achieving efficient homomorphic inference. The paper is available on arXiv under ID 2607.23478.
Key facts
- ATLAS automates polynomial approximation configuration for FHE transformer inference.
- Setup time reduced from days to one hour.
- FHE requires replacing softmax, normalization, and activation with polynomial approximations.
- Multiplicative depth of approximations dominates inference cost.
- Existing frameworks use uniform hyperparameters across all layers.
- Different layers tolerate different approximation error levels.
- ATLAS enables per-layer configuration to reduce latency.
- Paper available on arXiv: 2607.23478.
Entities
Institutions
- arXiv