Goten: GPU-Outsourcing Trusted Execution of Neural Network Training

Lucien K. L. Ng, Sherman S. M. Chow, Anna P. Y. Woo, Donald P. H. Wong, Yongjun ZhaoAAAI 2021 · DOI

Introduction

Training deep neural networks (DNNs) on sensitive data, such as medical records or genomic sequences, presents a significant privacy challenge. Current solutions generally fall into two categories: pure cryptographic methods like Multi-Party Computation (MPC), which are mathematically secure but computationally prohibitive, and Trusted Execution Environments (TEEs) like Intel SGX. While TEEs offer hardware-level security, they are often bottlenecked by weak CPUs and extremely limited secure memory (typically 90MB–128MB), making them ill-suited for the massive matrix operations required in modern deep learning.

Goten introduces a hybrid framework that bridges this gap by securely outsourcing heavy linear computations from TEEs to untrusted, high-performance GPUs. By combining a novel round-optimal protocol with dynamic quantization, Goten achieves a 132x speedup over state-of-the-art MPC methods and a 6.8x speedup over TEE-only baselines, all while maintaining the confidentiality of training data and model weights.

Methodology

Goten’s architecture is designed to treat the TEE as the "brain" and the GPU as the "muscle." The TEE handles sensitive non-linear layers (e.g., ReLU) and randomness generation, while the GPU performs heavy linear algebra on masked data.

Goten System Architecture: The framework partitions the training process, using secure enclaves for control and untrusted GPUs for heavy linear computation.

The framework relies on three core innovations:

  1. Secure GPU-Outsourcing Protocol: An adaptation of "Beaver’s Triplets" optimized for the TEE-GPU setting. This protocol is "round-optimal," minimizing the communication overhead between the CPU and GPU.
  2. Dynamic Quantization (SWALP): Since GPUs natively support floating-point arithmetic rather than the modular arithmetic used in cryptography, Goten employs Stochastic Weight Averaging in Low Precision (SWALP). This allows 53-bit integers to be represented within 64-bit double-precision floats without losing integrity.
  3. Memory-Aware Engineering: To overcome the 90MB memory limit of Intel SGX, Goten implements operation fusion and custom memory management to reduce data transfers across the secure enclave boundary.

Key Results

The researchers evaluated Goten using the CIFAR-10 image dataset and the Invasive Ductal Carcinoma (IDC) medical dataset. The results demonstrate that Goten significantly reduces the time required to reach high accuracy compared to previous secure training methods.

Training Performance: Goten reaches high test accuracy significantly faster than the TEE-only CaffeScone baseline.
  • Throughput: Goten outperformed the Falcon MPC framework by 132.64x on a VGG-11 model.
  • Efficiency: Compared to the TEE-only baseline (CaffeScone), Goten was 6.84x faster overall.
  • Accuracy: On CIFAR-10, Goten reached 89% accuracy, proving that its low-bitwidth quantization does not significantly degrade model performance. On the IDC medical dataset, it reached 82% accuracy in just 94 seconds—a 13.6x improvement over TEE-only approaches.

Discussion and Limitations

While Goten offers a massive leap in performance, it highlights that communication is the primary bottleneck in distributed private training. The system requires a minimum bandwidth of approximately 100Mbps to outperform TEE-only execution. Furthermore, while Goten is exceptionally fast at reaching high accuracy levels, it can encounter a "90% wall," where reaching the final percentage points of accuracy takes more epochs than non-quantized baselines.

Future work will focus on adapting Goten to other TEE architectures (like AMD SEV) and further optimizing the interface between secure enclaves and GPUs to reduce bandwidth sensitivity.

Summary updated

BibTeX
@inproceedings{ng2021goten,
  author = {Lucien K. L. Ng and Sherman S. M. Chow and Anna P. Y. Woo and Donald P. H. Wong and Yongjun Zhao},
  title = {{Goten: GPU-Outsourcing Trusted Execution of Neural Network Training}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i17.17746},
}
Figure