Heterogeneous CPU-GPU System Confidential Computing Survey

doi:10.3969/j.issn.1671-1122.2025.11.002

Abstract

Abstract:

With the widespread adoption of data-intensive applications such as artificial intelligence, heterogeneous computing systems centered on CPU and GPU have become essential infrastructure. However, in untrusted environments such as cloud and edge computing, sensitive data face severe security threats during processing, which cannot be effectively mitigated by traditional encryption methods. Confidential computing, leveraging hardware-based trusted execution environments (TEE), provides an effective mechanism for protecting data in use. Nevertheless, existing technologies have primarily focused on CPU. Extending TEE security boundaries seamlessly to GPU, the core of modern computing engines, has therefore become a major focus of both academic and industrial research. This paper provided a comprehensive review of confidential computing technologies in CPU-GPU heterogeneous systems. It first revisited the fundamental concepts of confidential computing and analyzed representative attack vectors targeting GPU. Subsequently, existing GPU confidential computing solutions were categorized into three paradigms: hardware-assisted, hardware-software co-design, and software-based approaches. Finally, the key challenges in this domain were summarized, and potential directions for future research were discussed.

Key words: confidential computing, trusted execution environment, heterogeneous computing, GPU

CLC Number:

TP309

HAO Meng, LI Jiayong, YANG Hongwei, ZHANG Weizhe. Heterogeneous CPU-GPU System Confidential Computing Survey[J]. Netinfo Security, 2025, 25(11): 1658-1672.

Figures/Tables 5

References 43

[1]	KHOKHAR A A, PRASANNA V K, SHAABAN M E, et al. Heterogeneous Computing: Challenges and Opportunities[J]. Computer, 2002, 26(6): 18-27.
[2]	MITTAL S, VETTER J S. A Survey of CPU-GPU Heterogeneous Computing Techniques[J]. ACM Computing Surveys (CSUR), 2015, 47(4): 1-35.
[3]	WANG Wei, HU Yin, CHEN Lianmu, et al. Accelerating Fully Homomorphic Encryption Using GPU[C]// IEEE. 2012 IEEE Conference on High Performance Extreme Computing. New York: IEEE, 2012: 1-5.
[4]	KNOTT B, VENKATARAMAN S, HANNUN A, et al. Crypten: Secure Multi-Party Computation Meets Machine Learning[J]. Advances in Neural Information Processing Systems, 2021, 34: 4961-4973.
[5]	VOLOS S, VASWANI K, BRUNO R. Graviton: Trusted Execution Environments on GPUs[C]// USENIX. The 13th USENIX Symposium on Operating Systems Design and Implementation. Berkeley: USENIX, 2018: 681-696.
[6]	COSTAN V, DEVADAS S. Intel SGX Explained[EB/OL]. (2017-02-21)[2025-06-30]. https://eprint.iacr.org/2016/086.pdf.
[7]	AMD. AMD Secure Encrypted Virtualization (SEV)[EB/OL]. (2020-01-01) [2025-06-30]. https://www.amd.com/en/developer/sev.html.
[8]	Confidential Computing Consortium. A Technical Analysis of Confidential Computing[EB/OL]. (2023-03-10)[2025-06-30]. https://confidentialcomputing.io/wp-content/uploads/sites/10/2023/03/CCC-A-Technical-Analysis-of-Confidential-Computing-v1.3_unlocked.pdf.
[9]	Trusted Computing Group. TCPA_Main_TCG_Architecture[EB/OL]. (2002-02-22)[2025-06-30]. https://trustedcomputinggroup.org/wp-content/uploads/TCPA_Main_TCG_Architecture_v1_1b.pdf.
[10]	JAMES G. Intel Trusted Execution Technology: Hardware-Based Technology for Enhancing Server Platform Security[EB/OL]. (2009-06-12)[2025-06-30]. https://www.intel.com/content/dam/www/public/us/en/documents/white-papers/trusted-execution-technology-security-paper.pdf.
[11]	MCKEEN F, ALEXANDROVICH I, BERENZON A, et al. Innovative Instructions and Software Model for Isolated Execution[EB/OL]. (2013-06-23) [2025-06-30]. https://doi.org/10.1145/2487726.2488368.
[12]	Intel. Intel^® 64 and IA-32 Architectures Software Developer’s Manual[EB/OL]. (2017-07-20)[2025-06-30]. https://ia600105.us.archive.org/6/items/Intel_64_and_IA-32_Architectures_Software_Developer_s_Manual_2017_07/325462-sdm-vol-1-2abcd-3abcd.pdf.
[13]	KAPLAN D, POWELL J, WOLLER T. AMD Memory Encryption[EB/OL]. (2021-10-18)[2025-06-30]. https://docs.amd.com/v/u/en-US/memory-encryption-white-paper.
[14]	AMD. AMD SEV-SNP: Strengthening VM Isolation with Integrity Protection and More[EB/OL]. (2020-01-01)[2025-06-30]. https://docs.amd.com/v/u/en-US/SEV-SNP-strengthening-vm-isolation-with-integrity-protection-and-more.
[15]	Intel. Intel^® Trust Domain Extensions[EB/OL]. (2021-05-01) [2025-06-30]. https://cdrdv2.intel.com/v1/dl/getContent/690419.
[16]	ARM. ARM Security Technology Building a Secure System Using TrustZone Technology[EB/OL]. (2008-12-01)[2025-06-30]. https://developer.arm.com/documentation/PRD29-GENC-009492/latest/.
[17]	PINTO S, SANTOS N. Demystifying ARM Trustzone: A Comprehensive Survey[J]. ACM Computing Surveys (CSUR), 2019, 51(6): 1-36.
[18]	ARM. ARM CCA Will Put Confidential Compute in the Hands of Every Developer[EB/OL]. (2021-06-23)[2025-06-30]. https://newsroom.arm.com/news/arm-cca-will-put-confidential-compute-in-the-hands-of-every-developer.
[19]	ARM. ARM Realm Management Extension(RME) System Architecture[EB/OL]. (2021-06-23)[2025-06-30]. https://developer.arm.com/documentation/den0129/latest/.
[20]	TF-HAFNIUM. Hafnium Is the Secure Partition Manager(SPM) Reference Implementation[EB/OL]. (2018-06-28)[2025-06-30]. https://github.com/TF-Hafnium/hafnium.
[21]	COSTAN V, LEBEDEV I, DEVADAS S. Sanctum: Minimal Hardware Extensions for Strong Software Isolation[C]// USENIX. The 25th USENIX Security Symposium. Berkeley: USENIX, 2016: 857-874.
[22]	LEE D, KOHLBRENNER D, SHINDE S, et al. Keystone: An Open Framework for Architecting Trusted Execution Environments[C]// ACM. The Fifteenth European Conference on Computer Systems. New York: ACM, 2020: 1-16.
[23]	NASAHL P, SCHILLING R, WERNER M, et al. HECTOR-V: A Heterogeneous CPU Architecture for a Secure RISC-V Execution Environment[C]// ACM. The 2021 ACM Asia Conference on Computer and Communications Security. New York: ACM, 2021: 187-199.
[24]	OZGA W, HUNT G D H, LE M V, et al. ACE: Confidential Computing for Embedded RISC-V Systems[EB/OL]. (2025-05-19) [2025-06-30]. https://arxiv.org/abs/2505.12995.
[25]	DUTTA S B, NAGHIBIJOUYBARI H, ABU-GHAZALEH N, et al. Leaky Buddies: Cross-Component Covert Channels on Integrated CPU-GPU Systems[C]// IEEE. 2021 ACM/IEEE 48th Annual International Symposium on Computer Architecture (ISCA). New York: IEEE, 2021: 972-984.
[26]	NAGHIBIJOUYBARI H, KHASAWNEH K N, ABU-GHAZALEH N. Constructing and Characterizing Covert Channels on GPGPUs[C]// ACM. The 50th Annual IEEE/ACM International Symposium on Microarchitecture. New York: ACM, 2017: 354-366.
[27]	AHN J, KIM J, KASAN H, et al. Network-on-Chip Microarchitecture-Based Covert Channel in GPUs[C]// ACM. The 54th Annual IEEE/ACM International Symposium on Microarchitecture. New York: ACM, 2021: 565-577.
[28]	DUTTA S B, NAGHIBIJOUYBARI H, GUPTA A, et al. Spy in the GPU-Box: Covert and Side Channel Attacks on Multi-GPU Systems[C]// ACM. The 50th Annual International Symposium on Computer Architecture. New York: ACM, 2023: 1-13.
[29]	ZHANG Zhenkai, ALLEN T, YAO Fan, et al. Tunnels for Bootlegging: Fully Reverse-Engineering GPU TLBs for Challenging Isolation Guarantees of NVIDIA MIG[C]// ACM. The 2023 ACM SIGSAC Conference on Computer and Communications Security. New York: ACM, 2023: 960-974.
[30]	ZHU Yuankun, CHENG Yueqiang, ZHOU Husheng, et al. Hermes Attack: Steal DNN Models With Lossless Inference Accuracy[C]// USENIX. The 30th USENIX Security Symposium. Berkeley: USENIX, 2021: 889-906.
[31]	HU Xing, LIANG Ling, LI Shuangchen, et al. Deepsniffer: A DNN Model Extraction Framework Based on Learning Architectural Hints[C]// ACM. The 25th International Conference on Architectural Support for Programming Languages and Operating Systems. New York: ACM, 2020: 385-399.
[32]	LUO Chao, FEI Yunsi, LUO Pei, et al. Side-Channel Power Analysis of a GPU AES Implementation[C]// IEEE. 2015 33rd IEEE International Conference on Computer Design (ICCD). New York: IEEE, 2015: 281-288.
[33]	GAO Yiwen, ZHANG Hailong, CHENG Wei, et al. Electro-Magnetic Analysis of GPU-Based AES Implementation[C]// ACM. The 55th Annual Design Automation Conference. New York: ACM, 2018: 1-6.
[34]	ZHAN Zihao, ZHANG Zhenkai, LIANG Sisheng, et al. Graphics Peeping Unit: Exploiting EM Side-Channel Information of GPUs to Eavesdrop on Your Neighbors[C]// IEEE. 2022 IEEE Symposium on Security and Privacy. New York: IEEE, 2022: 1440-1457.
[35]	JANG I, TANG A, KIM T, et al. Heterogeneous Isolated Execution for Commodity GPUs[C]// ACM. The 24th International Conference on Architectural Support for Programming Languages and Operating Systems. New York: ACM, 2019: 455-468.
[36]	DENG Yunjie, WANG Chenxu, YU Shunchang, et al. Strongbox: A GPU TEE on ARM Endpoints[C]// ACM. The 2022 ACM SIGSAC Conference on Computer and Communications Security. New York: ACM, 2022: 769-783.
[37]	WANG Chenxu, ZHANG Fengwei, DENG Yunjie, et al. CAGE: Complementing Arm CCA With GPU Extensions[EB/OL]. (2024-02-27)[2025-06-30]. https://www.ndss-symposium.org/wp-content/uploads/2024-763-paper.pdf.
[38]	MAI Haohui, ZHAO Jiacheng, ZHENG Hongren, et al. Honeycomb: Secure and Efficient GPU Executions via Static Validation[C]// USENIX. The 17th USENIX Symposium on Operating Systems Design and Implementation. Berkeley: USENIX, 2023: 155-172.
[39]	IVANOV A, ROTHENBERGER B, DETHISE A, et al. SAGE: Software-Based Attestation for GPU Execution[C]// USENIX. 2023 USENIX Annual Technical Conference. Berkeley: USENIX, 2023: 485-499.
[40]	ZHU Jianping, HOU Rui, WANG Xiaofeng, et al. Enabling Rack-Scale Confidential Computing Using Heterogeneous Trusted Execution Environment[C]// IEEE. 2020 IEEE Symposium on Security and Privacy (SP). New York: IEEE, 2020: 1450-1465.
[41]	NVIDIA. NVIDIA Confidential Computing[EB/OL]. (2022-04-25) [2025-06-30]. https://www.nvidia.com/en-us/data-center/solutions/confidential-computing/.
[42]	GU Zhongshu, VALDEZ E, AHMED S, et al. NVIDIA GPU Confidential Computing Demystified[EB/OL]. (2025-07-03)[2025-07-15]. https://arxiv.org/abs/2507.02770.
[43]	XU Tao, KONG Shuaidi, LIU Caihua, et al. A Review of Heterogeneous Confidential Computing[J]. Journal of Jilin University (Engineering Edition), 2025, 55(3): 755-770.
	徐涛, 孔帅迪, 刘才华, 等. 异构机密计算综述[J]. 吉林大学学报(工学版), 2025, 55(3):755-770.

架构	技术	发布年份	隔离模型	核心硬件/机制
x86	TCG TPM 1.1b	2003	静态测量信任根	PCRs，CRTM，加密协处理器
	Intel TXT	2006	动态测量信任根	GETSEC指令，SINIT ACM，动态PCRs
	Intel SGX	2015	应用程序飞地	ENCLU/ENCLS指令，EPC，EPCM
	AMD SEV	2017	机密虚拟机	AMD-SP，内存控制器AES引擎
	AMD SEV-SNP	2020	机密虚拟机	SEV机制 + RMP
	Intel TDX	2021	机密虚拟机	VMX，MKTME，SEAM模式，TDX模块
ARM	TrustZone	2003	系统分区	NS-bit，安全监视器，TZASC/TZPC
ARM	CCA / RME	2020	机密虚拟机	Realm/Root世界，GPT，RMM固件
RISC-V	Sanctum	2016	应用程序飞地	PMP
	Keystone	2020	应用程序飞地（框架）	PMP，可定制的安全监视器（SM）和运行时（RT）
	HECTOR-V	2021	协处理器	异构多核，专用的安全协处理器
	ACE	2025	机密虚拟机	H-extension，PMP，形式化验证的TSM

	硬件依赖	性能开销（典型）	保护方法
Graviton	高（需改GPU）	17%~33%	修改GPU命令处理器；PMM
HIX StrongBox	低（需改CPU/平台）无（利用现有ARM特性）	平均26% 4.7%~15.3%	CPU侧IOMMU/PCIe扩展；GPU飞地；ARM TrustZone；Stage-2地址翻译
CAGE	无（利用现有ARM特性）	平均2.45%	ARM CCA；影子任务；双向隔离
HETEE	高（需专用HETEE Box）	小于2.2%	机架级PCIe物理隔离
Honeycomb	无（需CPU TEE）	2%（大模型）	静态二进制验证；双重监控器
SAGE	无（需CPU TEE）	验证阶段开销；执行时无损	基于时间的执行证明
H100	高（专用GPU）	—	硬件防火墙；端到端PCIe加密