Secure computation techniques like Fully Homomorphic Encryption (FHE), Secure Multi-Party Computation (MPC), and Zero-Knowledge Proofs (ZKP) offer a powerful promise: the ability to process sensitive data without ever seeing it. However, these methods have historically been sidelined by a massive "tax" on performance. Early implementations were often described as theoretical "monstrosities" due to the overwhelming computational overhead of modular arithmetic and encryption Setty et al..
The path to practicality lies in a fundamental realization: while cryptographic operations are slow, they are also highly parallelizable. By offloading these workloads to Graphics Processing Units (GPUs) and specialized hardware, researchers are transforming secure computation from a theoretical curiosity into a viable tool for real-time applications.
The Mathematical Bottlenecks
The primary reason secure computation is slow is that it replaces simple machine instructions with complex mathematical transformations. Three specific operations dominate the runtime:
- Number Theoretic Transform (NTT): Essential for polynomial multiplication in FHE and ZKPs. Wooseok Choi et al. note that NTT often accounts for the majority of computational runtime in these systems.
- Multi-Scalar Multiplication (MSM): A core bottleneck in generating ZKPs.
- Modular Exponentiation and XORs: The building blocks of bilinear pairings and private information retrieval.
These operations are computationally expensive but "data-parallel"—the same operation is performed on thousands of different data points simultaneously. This makes them a perfect fit for the Single Instruction, Multiple Data (SIMD) architecture of GPUs. For instance, Raymond Cheng et al. achieved a 20-fold increase in processing speed for private messaging by offloading large-scale XOR operations to the GPU.
Mapping Cryptography to Hardware
To achieve these speedups, researchers do not just port code; they reimagine the underlying algorithms to fit hardware constraints.
Shengyu Fan demonstrates this by transforming the NTT bottleneck into a matrix-multiplication problem. This allows the use of NVIDIA’s Tensor Core Units (TCUs)—hardware specifically optimized for deep learning—to perform cryptographic arithmetic. Similarly, Tao Lu et al. reimagined Multi-Scalar Multiplication as a sparse matrix problem, allowing the GPU to handle all major computational stages of a proof and avoiding the performance limits of Amdahl's Law.
In the realm of MPC, the challenge is often the irregular logic of standard algorithms. GTree addresses this by redesigning the branching logic of decision trees into regular, array-based operations. This "GPU-friendly" structure allows for a 126× speedup in secure inference by ensuring the hardware cores are never left idle by complex conditional paths.
Overcoming the Memory and Noise Barriers
Even with massive parallelization, two major hurdles remain: memory bandwidth and "noise" in FHE.
FHE ciphertexts are significantly larger than the data they hide, leading to massive data movement. Jaiyoung Park et al. address this through memory-centric kernel fusion, which reduces the need to move data between the GPU's fast on-chip memory and its slower main memory.
Furthermore, FHE computations accumulate "noise" that must be periodically cleared through a process called bootstrapping. This is notoriously slow on CPUs. Leo de Castro et al. showed that accelerating this bootstrapping phase on GPUs can speed up private Large Language Model (LLM) inference by over 200x, turning a process that took hours into one that takes minutes.
Toward Secure-by-Design AI
The ultimate goal of this hardware acceleration is to enable private machine learning at scale. Frameworks like CrypTen Knott et al. and AriaNN Ryffel et al. are bridging the gap by providing PyTorch-like APIs that automatically offload cryptographic tasks to GPUs.
In some cases, the neural networks themselves are being redesigned to be more "hardware-cryptography friendly." Jiajie Li et al. proposed replacing standard ReLU activations with square functions ($x^2$). Because $x^2$ is a simple arithmetic operation, it can be parallelized much more efficiently than the comparison-based logic of ReLU, leading to speedups of up to 700x on GPUs.
By co-designing cryptographic protocols, software frameworks, and hardware architectures, the research community is moving toward a future where privacy is not a performance trade-off, but a standard feature of high-performance computing.
Key Foundational Work
Acceleration of Composite Order Bilinear Pairing on Graphics Hardware (2012) — contributed 2 concept elements to this direction.
Tsunami: massively parallel homomorphic hashing on many-core GPUs (2012) — contributed 2 concept elements to this direction.
Accelerating NTRU based Homomorphic Encryption using GPUs (2014) — contributed 1 concept elements to this direction.