Secured and Privacy-Preserving GPU-Based Machine Learning Inference in Trusted Execution Environment: A Comprehensive Survey

• Published in: International Conference on Communication Systems and Networks (Online) pp. 207 - 216
• Main Authors: Chaudhuri, Arunava, Shukla, Shubhi, Bhattacharya, Sarani, Mukhopadhyay, Debdeep
• Format: Conference Proceeding
• Language: English
• Published: IEEE 06.01.2025
• Subjects: Data models; Data privacy; Deep learning; GPU; Hardware acceleration; Privacy; Security; Stability analysis; Surveys; Training; Training data; trusted execution environment
• ISSN: 2155-2509
• DOI: 10.1109/COMSNETS63942.2025.10885734

With the rapid advancement of machine learning (ML) models and their widespread application across various sectors such as intrusion detection, medical diagnosis, natural language processing, and autonomous driving, these technologies have achieved remarkable success. However, this progress has also raised significant concerns about ensuring the security of ML models and protecting both private training data and model outputs from getting exposed in a shared cloud environment. To address these challenges, researchers have proposed various methodologies to create privacy-preserving, secure, and trustworthy model execution environments to prevent adversarial attacks. This study provides a comprehensive review of Trusted Execution Environment (TEE) implementations across different hardware accelerators. It also offers an overview of modern techniques for preserving privacy and security in execution environments, while identifying critical research gaps that require attention. In essence, this survey is an important resource for researchers, providing insights into recent methodologies and guiding them to focus on pressing research challenges.