New AI Design Slashes Privacy Computation Times by up to 9.78x

Researchers are tackling the significant challenge of computational overhead in privacy-preserving deep learning, a field crucial for secure Machine Learning as a Service. Yifei Cai from Iowa State University, Yizhou Feng from Old Dominion University, and Qiao Zhang from Shandong University, alongside Chunsheng Xin and Hongyi Wu et al., present a novel approach that moves beyond adapting existing neural network architectures for homomorphic encryption (HE). Their work demonstrates that substantial efficiency gains require networks specifically designed with HE in mind, introducing StriaNet which incorporates the innovative StriaBlock to minimise computationally expensive rotation operations. By employing focused constraints and channel packing-aware scaling, the team achieves speedups of up to 9.78x on ImageNet, representing a considerable advancement towards practical, fast, and secure inference with HE.

StrNet architecture accelerates homomorphic encryption via rotation-reducing design principles and efficient layer fusion

Researchers have developed StriaNet, a novel deep learning architecture specifically designed to accelerate privacy-preserving machine learning using homomorphic encryption. The work addresses a critical bottleneck in Machine Learning as a Service, where clients are hesitant to share private data and servers need to protect model parameters.
By tailoring network design to the unique computational demands of homomorphic encryption, rather than adapting existing plaintext models, substantial efficiency gains have been achieved. StriaNet introduces a building block, termed StriaBlock, that dramatically reduces the computational cost of rotation, the most expensive operation within homomorphic encryption.

StriaBlock integrates ExRot-Free Convolution and a novel Cross Kernel, completely eliminating external rotations and requiring only 19% of the internal rotations used by conventional plaintext models. This innovative design is coupled with two architectural principles: the Focused Constraint Principle, which prioritises cost reduction in critical areas while maintaining flexibility elsewhere, and the Channel Packing-Aware Scaling Principle, which dynamically adjusts network parameters to optimise ciphertext channel capacity at varying depths.
Together, these strategies enable 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 varying scales, including ImageNet, Tiny ImageNet, and CIFAR-10. Evaluations demonstrate that StriaNet achieves significant speedups at comparable accuracy levels.

Specifically, the research reports speedups of 9.78x on ImageNet, 6.01x on Tiny ImageNet, and 9.24x on CIFAR-10. These improvements surpass those achieved by optimising existing models, highlighting the benefits of an architecture designed specifically for homomorphic encryption. This work moves beyond small-scale dataset evaluations, providing extensive results across large, medium, and small benchmarks to demonstrate broad applicability and impact. The development of StriaNet represents a significant step towards practical, real-time privacy-preserving machine learning applications.

Mitigating Rotational Complexity in Homomorphic Encryption via StriaBlock and Architectural Optimisation yields significant performance gains

A detailed analysis of rotation costs within homomorphic encryption (HE) operations motivated the development of StriaNet, a neural network architecture specifically tailored for privacy-preserving deep learning. The research began by identifying that rotation constitutes the most computationally expensive operation in HE, encompassing both internal and external rotation processes.

To address this, researchers designed StriaBlock, a novel building block integrating ExRot-Free Convolution and a Cross Kernel, effectively eliminating external rotations and reducing internal rotations to 19% of those required by standard plaintext models. This reduction in rotational complexity forms the core of the efficiency gains observed.

Beyond the building block, the study implemented two architectural principles to further optimise HE cost. The Focused Constraint Principle strategically limits cost-sensitive factors within the network while maintaining flexibility in other areas. Simultaneously, the Channel Packing-Aware Scaling Principle adapts bottleneck ratios to the ciphertext channel capacity, which naturally varies with network depth.

These principles collaboratively manage both local computational costs and overall end-to-end HE expense, resulting in a balanced network design. Evaluation of StriaNet involved comprehensive testing across datasets of varying scales, including ImageNet, Tiny ImageNet, and CIFAR-10. Performance was measured by comparing StriaNet’s speed against existing models at comparable accuracy levels.

Results demonstrated speedups of 9.78x on ImageNet, 6.01x on Tiny ImageNet, and 9.24x on CIFAR-10. Furthermore, the efficiency of StriaNet was shown to scale positively 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 optimising models initially created for plaintext inference.

Accelerated Homomorphic Encryption via Rotation-Reducing Network Design enables efficient and practical privacy-preserving computation

StriaNet, a highly HE-efficient deep learning network, achieves speedups of 9.78x on ImageNet, 6.01x on Tiny ImageNet, and 9.24x on CIFAR-10 at comparable accuracy levels. This performance is enabled by a novel building block, StriaBlock, and associated architectural principles designed specifically for homomorphic encryption (HE).

StriaBlock targets the most computationally expensive HE operation, rotation, integrating ExRot-Free Convolution and a novel Cross Kernel to eliminate external rotations. Internal rotations are reduced to 19% of those used in standard plaintext models, significantly decreasing computational load. The research introduces two key architectural principles: the Focused Constraint Principle and the Channel Packing-Aware Scaling Principle.

The Focused Constraint Principle limits cost-sensitive factors while maintaining flexibility, and the Channel Packing-Aware Scaling Principle adapts bottleneck ratios to ciphertext channel capacity, which varies with network depth. These strategies collectively control both local and end-to-end HE cost, resulting in a balanced network tailored for HE operations.

A 31.4x reduction in the number of l layer operations was observed, leading to a 13.1x speedup for the target layer and a 1.73x overall speedup in end-to-end inference. The study focuses on privacy-preserving Machine Learning as a Service (PP MLaaS) utilizing HE for linear computations. Linear operations comprise over 90% of the total inference cost in deep learning models, with convolution accounting for approximately 99.05% of the linear computation in ResNet-50.

StriaNet is compatible with existing PP MLaaS frameworks and optimizations, serving as a strong baseline model for HE-based scenarios by maximizing the efficiency of these optimizations. The network maintains the same level of security as existing approaches, introducing no additional computational modules or algorithms.

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

Optimised StriaBlock design for rotation reduction in encrypted deep learning improves model accuracy and security

StriaNet, a novel neural network architecture, substantially 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 research demonstrates that networks specifically tailored for homomorphic encryption offer significant performance gains. The core of this advancement lies in the StriaBlock building block, which targets the computationally intensive rotation operation inherent in homomorphic encryption. By integrating ExRot-Free Convolution and a Cross Kernel, StriaBlock eliminates external rotations and reduces internal rotation usage to 19% of that required by conventional plaintext models.

Furthermore, the architecture employs two guiding principles: the Focused Constraint Principle, which prioritises cost reduction in critical areas, and the Channel Packing-Aware Scaling Principle, which optimises network structure based on ciphertext channel capacity. Evaluations across ImageNet, Tiny ImageNet, and CIFAR-10 datasets reveal speedups of 9.78x, 6.01x, and 9.24x, respectively, while maintaining comparable accuracy.

This work addresses a key limitation in privacy-preserving machine learning as a service by moving beyond adaptation of existing models towards a design explicitly suited to the constraints of homomorphic encryption. The resulting efficiency improvements facilitate more practical deployment of privacy-preserving deep learning systems.

The authors acknowledge that the current work does not explore neural architecture search for further optimisation. Future research will focus on incorporating these techniques to potentially enhance StriaNet’s performance still further.