New AI design reduces privacy computation time by up to 9.78x

Researchers are tackling the critical challenge of computational overhead in privacy-preserving deep learning, a critical area for secure machine learning as a service. Yifei Cai of Iowa State University, Yizhou Feng of Old Dominion University, and Qiao Zhang of Shandong University, along with Chunsheng Xin, Hongyi Wu, and others present a new approach that goes beyond adapting existing neural network architectures for homomorphic encryption (HE). Their work demonstrates that significantly improved efficiency requires networks specifically designed with HE in mind, and they introduce StriaNet, which incorporates an innovative StriaBlock that minimizes computationally expensive rotation operations. By employing focused constraints and channel-packing-aware scaling, the team achieved up to 9.78x speedup on ImageNet. This represents significant progress toward practical, fast, and secure inference with HE.

The StrNet architecture accelerates homomorphic encryption through design principles that reduce rotation and efficient layer fusion.

Researchers have developed StriaNet, a new deep learning architecture specifically designed to accelerate privacy-preserving machine learning using homomorphic encryption. This effort addresses a critical bottleneck in machine learning as a service where clients are reluctant to share private data and servers need to protect model parameters.
Significant efficiency gains were achieved by tailoring the network design to the unique computational demands of homomorphic encryption, rather than adapting existing plaintext models. StriaNet introduces a building block called StriaBlock that significantly reduces the computational cost of rotation, the most expensive operation within homomorphic encryption.

StriaBlock integrates ExRot-Free Convolution with a new cross kernel, completely eliminating external rotations and requiring only 19% of the internal rotations used in traditional plaintext models. This innovative design combines two architectural principles: One is a centralized constraint principle that prioritizes cost reduction in critical areas while maintaining flexibility elsewhere, and the other is a channel packing-aware scaling principle that dynamically adjusts network parameters to optimize the capacity of ciphertext channels at different depths.
Combining these strategies enables a balanced network architecture that controls both local and end-to-end homomorphic encryption costs. The resulting network, StriaNet, has undergone comprehensive evaluation across datasets of various sizes, including ImageNet, Tiny ImageNet, and CIFAR-10. Evaluations demonstrate that StriaNet achieves significant speedups at comparable accuracy levels.

Specifically, the study reports speedups of 9.78x for ImageNet, 6.01x for Tiny ImageNet, and 9.24x for CIFAR-10. These improvements exceed those achieved by optimizing existing models and highlight the benefits of an architecture designed specifically for homomorphic encryption. This study goes beyond a small dataset evaluation to provide extensive results across large, medium, and small benchmarks, demonstrating broad applicability and impact. The development of StriaNet represents an important step toward practical real-time, privacy-preserving machine learning applications.

Reducing the rotational complexity of homomorphic encryption through StriaBlock and architectural optimizations significantly improves performance

A detailed analysis of rotation costs within homomorphic encryption (HE) operations led to the development of StriaNet, a privacy-preserving deep learning-specific neural network architecture. The study began by identifying that rotation is the most computationally expensive operation in HE, including both internal and external rotation processes.

To address this, researchers designed StriaBlock, a new building block that integrates ExRot-Free Convolution and Cross Kernel. This effectively eliminates external rotations and reduces internal rotations to 19% of what is required by the standard plaintext model. This reduction in rotational complexity forms the core of the observed efficiency gains.

In this study, two architectural principles were implemented to further optimize HE costs beyond the building blocks. The centralization constraint principle strategically limits cost-sensitive elements in the network while maintaining flexibility in other areas. At the same time, the channel packing-aware scaling principle adapts the bottleneck ratio to the capacity of the ciphertext channel. The capacity of the ciphertext channel naturally varies with the depth of the network.

These principles work together to manage both local computational costs and overall end-to-end HE costs, resulting in a balanced network design. Our evaluation of StriaNet includes comprehensive testing across datasets of various sizes, including ImageNet, Tiny ImageNet, and CIFAR-10. Performance was measured by comparing the speed of StriaNet to existing models at comparable accuracy levels.

The results showed a speedup of 9.78x on ImageNet, 6.01x on Tiny ImageNet, and 9.24x on CIFAR-10. Additionally, we show that StriaNet’s efficiency increases with batch size, achieving up to a 5.2x speedup with a batch size of 512. These findings highlight the benefits of designing architectures specifically for HE, rather than optimizing models originally created for plaintext inference.

Fast homomorphic encryption with rotation-reduced network design enables efficient and practical privacy-preserving computation

StriaNet, a HE-efficient deep learning network, achieves 9.78x speedup on ImageNet, 6.01x on Tiny ImageNet, and 9.24x on CIFAR-10 at comparable accuracy levels. This performance is achieved through a new building block, StriaBlock, and related architectural principles designed specifically for homomorphic encryption (HE).

StriaBlock targets the most computationally intensive HE operation, rotation, and integrates ExRot-Free Convolution and a new cross kernel that eliminates external rotations. The internal rotation is reduced to 19% of that used in the standard plaintext model, significantly reducing the computational load. This study introduces two key architectural principles: the centralization constraint principle and the channel packing-aware scaling principle.

The centralization constraint principle limits cost-sensitive elements while maintaining flexibility, and the channel packing-aware scaling principle adapts the bottleneck ratio to the capacity of the ciphertext channel, which varies with network depth. These strategies collectively control both local and end-to-end HE costs, resulting in a balanced network tailored for HE operations.

We observed a 31.4x reduction in the number of l layer operations, which resulted in a 13.1x speedup for the target layer and a 1.73x overall speedup for end-to-end inference. This research focuses on privacy-preserving machine learning as a service (PP MLaaS) leveraging HE for linear computation. Linear operations account for more than 90% of the total inference cost in deep learning models, and convolutions account for approximately 99.05% of linear computations in ResNet-50.

StriaNet is compatible with existing PP MLaaS frameworks and optimizations, and serves as a powerful baseline model for HE-based scenarios by maximizing the efficiency of these optimizations. The network maintains the same level of security as existing approaches and does not introduce additional computational modules or algorithms.

Packed HE is utilized, where the client encrypts sensitive data and sends the ciphertext to the server for computation. Three basic HE operations are used: homomorphic addition, multiplication, and rotation, with rotation being the most computationally expensive. StriaNet’s design minimizes the need for rotation, increasing overall efficiency and enabling practical, privacy-preserving deep learning.

StriaBlock design optimized for rotation reduction in cryptographic deep learning improves model accuracy and security

StriaNet, a new neural network architecture, significantly improves the efficiency of privacy-preserving deep learning using homomorphic encryption. Current approaches often adapt existing networks designed for plaintext data, resulting in architectural inefficiencies when applied to encrypted data.

This study demonstrates that a network dedicated to homomorphic encryption provides significant performance improvements. The core of this advancement lies in the StriaBlock building block, which targets the computationally intensive rotation operations inherent in homomorphic encryption. By integrating ExRot-Free Convolution and Cross Kernel, StriaBlock eliminates external rotations and reduces internal rotation usage to 19% of that required by traditional plaintext models.

In addition, the architecture employs two guiding principles: the centralization constraint principle, which prioritizes cost reduction in critical areas, and the channel packing-aware scaling principle, which optimizes the network structure based on the capacity of the ciphertext channel. Evaluation across ImageNet, Tiny ImageNet, and CIFAR-10 datasets reveals speedups of 9.78x, 6.01x, and 9.24x, respectively, while maintaining comparable accuracy.

This work addresses a key limitation of privacy-preserving machine learning as a service by moving beyond adapting existing models to a design explicitly suited to the constraints of homomorphic encryption. The result is greater efficiency and facilitates more practical deployment of privacy-preserving deep learning systems.

The authors acknowledge that the current study does not consider exploring neural architectures for further optimization. Future research will focus on incorporating these techniques to further improve the performance of StriaNet.