Fixed Point Arithmetic (FPA) is widely used in Privacy-Preserving Machine Learning (PPML) to efficiently handle decimal values. However, repeated multiplications in FPA can lead to overflow, as the fractional part doubles in size with each multiplication. To address this, truncation is applied post-multiplication to maintain precision. Various truncation schemes based on Secure Multiparty Computation (MPC) exist, but trade-offs between accuracy and efficiency in PPML models and datasets remain underexplored. In this work, we analyze and consolidate different truncation approaches from the MPC literature. We conduct the first large-scale systematic evaluation of PPML inference accuracy across truncation schemes, ring sizes, neural network architectures, and datasets. Our study provides clear guidelines for selecting the optimal truncation scheme and parameters for PPML inference. All evaluations are implemented in the open-source HPMPC MPC framework, facilitating future research and adoption. Beyond our large scale evaluation, we also present improved constructions for each truncation scheme, achieving up to a fourfold reduction in communication and round complexity over existing schemes. Additionally, we introduce optimizations tailored for PPML, such as strategically fusing different neural network layers. This leads to a mixed-truncation scheme that balances truncation costs with accuracy, eliminating communication overhead in the online phase while matching the accuracy of plaintext floating-point PyTorch inference for VGG-16 on the ImageNet dataset.

In this work, we present novel protocols over rings for semi-honest secure three-party computation (3PC) and malicious four-party computation (4PC) with one corruption. While most existing works focus on improving total communication complexity, challenges such as network heterogeneity and computational complexity, which impact MPC performance in practice, remain underexplored. Our protocols address these issues by tolerating multiple arbitrarily weak network links between parties without any substantial decrease in performance. Additionally, they significantly reduce computational complexity by requiring up to half the number of basic instructions per gate compared to related work. These improvements lead to up to twice the throughput of state-of-the-art protocols in homogeneous network settings and up to eight times higher throughput in real-world heterogeneous settings. These advantages come at no additional cost: Our protocols maintain the best-known total communication complexity per multiplication, requiring 3 elements for 3PC and 5 elements for 4PC.We implemented our protocols alongside several state-of-the-art protocols (Replicated 3PC, ASTRA, Fantastic Four, Tetrad) in a novel open-source C++ framework optimized for high throughput. Five out of six implemented 3PC and 4PC protocols achieve more than one billion 32-bit multiplications or over 32 billion AND gates per second using our implementation in a 25 Gbit/s LAN environment. This represents the highest throughput achieved in 3PC and 4PC so far, outperforming existing frameworks like MP-SPDZ, ABY3, MPyC, and MOTION by two to three orders of magnitude.

Privacy-Preserving Machine Learning (PPML) is one of the most relevant use cases for Secure Multiparty Computation (MPC). While private training of large neural networks such as VGG-16 or ResNet-50 on state-of-the-art datasets such as ImageNet is still out of reach, given the performance overhead of MPC, GPU-based MPC frameworks are starting to achieve practical runtimes for private inference. However, we show that, unlike plaintext machine learning, using GPU acceleration for both linear (e.g., convolutions) and non-linear neural network layers (e.g., ReLU) is actually counterproductive in PPML. While GPUs effectively accelerate linear layers compared to CPU-based MPC implementations, the MPC circuits required to evaluate non-linear layers introduce memory overhead and frequent data movement between the GPU and the CPU to handle network communication. This results in slow ReLU performance and high GPU memory requirements in state-of-the-art GPU-based PPML frameworks, hindering them from scaling to multiple images per second inference throughput and more than eight images per batch on ImageNet. To overcome these limitations, we propose PIGEON, an open-source framework for Private Inference of Neural Networks. PIGEON employs a novel ABG programming model that switches between Arithmetic Vectorization and Bitslicing on the CPU for non-linear layers depending on the MPC-specific computation required while offloading linear layers to the GPU. Compared to the state-of-the-art PPML framework Piranha, PIGEON improves ReLU throughput by two orders of magnitude, reduces peak GPU memory utilization by one order of magnitude, and scales better with large batch sizes. This translates to one to two orders of magnitude improvements in throughput for large ImageNet batch sizes (e.g., 192) and more than 70% saturation of a 25 Gbit/s network.

Privacy-enhancing technologies (PETs) are becoming increasingly crucial for addressing customer needs, security, privacy (e. g., enhancing anonymity and confidentiality), and regulatory requirements. However, applying PETs in organizations requires a precise understanding of use cases, technologies, and limitations. This paper investigates several industrial use cases, their characteristics, and the potential applicability of PETs to these. We conduct expert interviews to identify and classify uses cases, a gray literature review of relevant open-source PET tools, and discuss how the use case characteristics can be addressed using PETs’ capabilities. While we focus mainly on automotive use cases, the results also apply to other use case domains.