Secure Floating-Point Training

Secure 2-party computation (2PC) of floating-point arithmetic is improving in performance and recent work runs deep learning algorithms with it, while being as numerically precise as commonly used machine learning (ML) frameworks like PyTorch. We find that the existing 2PC libraries for floatingpoint support generic computations and lack specialized support for ML training. Hence, their latency and communication costs for compound operations (e.g., dot products) are high. We provide novel specialized 2PC protocols for compound operations and prove their precision using numerical analysis. Our implementation BEACON outperforms state-of-the-art libraries for 2PC of floating-point by over 6×. We provide novel specialized protocols for compound operations occurring in ML training that are as precise as SECFLOAT-based protocols while being much more efficient. Note that, apart from SECFLOAT, all other prior works in secure training use approximations that lack formal precision guarantees [16, 17, 41, 43, 54, 55, 64, 66, 67] . Among these, KS22 [41] is the state-of-the-art whose approximations have high efficiency and have been shown to empirically match the end-to-end training accuracy provided by floating-point training on MNIST [46]/CIFAR [44] datasets. We show that the latency overheads of our provably precise protocols are < 6× over KS22. In particular,