Treffer: MLFormer: a high performance MPC linear inference framework for transformers.

Transformer-based models are widely used in natural language processing tasks, and their application has been further extended to computer vision as well. In their usage, data security has become a crucial concern when deploying deep learning services on cloud platforms. To address these security concerns, Multi-party computation (MPC) is employed to prevent data and model leakage during the inference process. However, Transformer model introduces several challenges for MPC computation, including the time overhead of the Softmax (normalized exponential) function, the accuracy issue caused by the “dynamic range” of approximated division and exponential, and the high memory overhead when processing long sequences. To overcome these challenges, we propose MLformer, an MPC-based inference framework for transformer models based on Crypten Knott et al. (Adv Neural Inf Process Syst 34: 4961–4973, 2021), a secure machine learning framework suggested by Facebook AI Research group, in the semi-honest adversary model. In this framework, we replace the softmax attention with linear attention, which has linear time and memory complexity with input length. The modification eliminates the softmax function entirely, resulting in lower time and memory overhead. To ensure the accuracy of linear attention, we propose the scaled linear attention to address the dynamic range issue caused by the MPC division used and a new approximate division function is proposed to reduce the computational time of the attention block. Furthermore, to improve the efficiency and accuracy of MPC exponential and reciprocal which are commonly used in transformer model, we propose a novel MPC exponential protocol and first integrate the efficient reciprocal protocol Bar-Ilan and Beaver (in Proceedings of the 8th annual ACM symposium on principles of distributed computing, pp. 201–209, 1989) to our framework. Additionally, we optimize the computation of causal linear attention, which is utilized in private inference of auto-regression tasks, using our novel CUDA kernel functions. All the proceeding optimizations contribute to the construction of a more accurate and efficient framework. The experimental results demonstrate that our framework achieves comparable accuracy with reduced inference time and GPU memory overhead compared to the original transformer model. The speedup reaches 78.79% compared to traditional private transformer with input length of 1024 patches. [ABSTRACT FROM AUTHOR]

Copyright of Journal of Cryptographic Engineering is the property of Springer Nature and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)