Andes Y. L. Kei (Chinese University of Hong Kong), Sherman S. M. Chow (Chinese University of Hong Kong)

Adoption of transformer-based machine learning models is growing, raising concerns about sensitive data exposure. Nonetheless, current secure inference solutions incur substantial overhead due to their extensive reliance on non-linear protocols, such as softmax and Gaussian error linear unit (GELU). Driven by numerical stability needs, softmax approximations (e.g., NeurIPS 2021) typically extract the maximum element of an input vector, incurring logarithmic rounds (in the input length). Existing GELU protocols (e.g., S&P 2024) use piecewise approximations with high-degree polynomials that rely heavily on secure multiplications and comparisons, which are expensive. Such complexities also hinder model owners who are not familiar with cryptography from easily deploying their custom models.

SHAFT, our proposed system, provides a secure, handy, accurate, and fast transformer inference framework for deployment. Highlights of our contributions include 1) the first constant-round softmax protocol for transformers, uniquely combining the benefits of input clipping and characteristics of ordinary differential equations, and 2) a highly accurate GELU protocol on a novel characterization designed for Fourier series approximation. Extending to broader contexts, our new protocols also apply to general neural networks using softmax as the final layer and to transformer architectures with different activation functions. Remarkably, SHAFT outperforms state-of-the-art SIGMA (PETS 2024), based on secret sharing, and BumbleBee (NDSS 2025), which additionally uses RLWE-based homomorphic encryption. More specifically, SHAFT minimizes communication by 25-41%. and matches SIGMA's running time while surpassing BumbleBee in running time by 4.6-5.3× on LANs and 2.9-4.4× on WANs. Alongside these improvements, SHAFT attains accuracy comparable to plaintext, confirming its numerical stability and accuracy. Next in this progression, SHAFT provides an accessible open-source framework for secure and handy deployment by smoothly integrating with the Hugging Face library (EMNLP Demos 2020).

View More Papers

Incorporating Gradients to Rules: Towards Lightweight, Adaptive Provenance-based Intrusion...

Lingzhi Wang (Northwestern University), Xiangmin Shen (Northwestern University), Weijian Li (Northwestern University), Zhenyuan LI (Zhejiang University), R. Sekar (Stony Brook University), Han Liu (Northwestern University), Yan Chen (Northwestern University)

Read More

RAIFLE: Reconstruction Attacks on Interaction-based Federated Learning with Adversarial...

Dzung Pham (University of Massachusetts Amherst), Shreyas Kulkarni (University of Massachusetts Amherst), Amir Houmansadr (University of Massachusetts Amherst)

Read More

Lend Me Your Beam: Privacy Implications of Plaintext Beamforming...

Rui Xiao (Zhejiang University), Xiankai Chen (Zhejiang University), Yinghui He (Nanyang Technological University), Jun Han (KAIST), Jinsong Han (Zhejiang University)

Read More

Explanation as a Watermark: Towards Harmless and Multi-bit Model...

Shuo Shao (Zhejiang University), Yiming Li (Zhejiang University), Hongwei Yao (Zhejiang University), Yiling He (Zhejiang University), Zhan Qin (Zhejiang University), Kui Ren (Zhejiang University)

Read More