Thursday 06 March 2025
Homomorphic encryption, a cryptographic technique that enables computations on encrypted data without decrypting it first, has long been touted as a solution for secure and private data processing in the cloud. However, its practical application has been hindered by the need for powerful hardware accelerators to speed up the computationally intensive operations involved.
Researchers have now developed a novel approach that leverages existing artificial intelligence (AI) chips to accelerate homomorphic encryption, making it more feasible for widespread adoption. The technique, dubbed CROSS, converts homomorphic encryption primitives into AI operators and optimizes them for execution on tensor processing units (TPUs), which are designed for machine learning workloads.
The key innovation lies in the development of a dedicated mapping for 28-bit moduli, a crucial component of homomorphic encryption. This involves a two-level precision breakdown: RNS decomposition, which breaks down large coefficients into smaller limbs, and chunk decomposition, which further divides these limbs into manageable chunks.
By optimizing this process for TPUs, CROSS achieves significant performance improvements compared to traditional CPU-based implementations. In fact, the authors demonstrate a 161-fold speedup on a Google TPUv4 compared to previous works on many-core CPUs and NVIDIA V100 GPUs.
The implications of this breakthrough are substantial. With CROSS, cloud-based services can now provide secure and private data processing without sacrificing performance. This enables a wide range of applications, from encrypted machine learning model serving to secure data analytics in the cloud.
One potential use case is in healthcare, where sensitive patient data could be processed on remote servers while maintaining confidentiality. Another area is in finance, where financial institutions could utilize CROSS-enabled encryption for secure data processing and analysis.
The authors’ approach has also shed light on the potential of repurposing AI accelerators for cryptographic applications. As TPUs become increasingly ubiquitous, this development could pave the way for a new generation of secure computing architectures.
Moreover, CROSS has sparked interest in exploring other areas where homomorphic encryption can be applied, such as private machine learning and secure data sharing. With its potential to revolutionize the field of cryptography, CROSS is an exciting development that warrants further exploration and research.
Cite this article: “Accelerating Homomorphic Encryption with AI Chips”, The Science Archive, 2025.
Homomorphic Encryption, Artificial Intelligence, Tpus, Machine Learning, Cloud Computing, Secure Data Processing, Cryptography, Ai Accelerators, Private Data Sharing, Secure Computing Architectures
