Specialized AI Development
We develop specialized AI for generating parallel confidential computing code and optimizing it to reduce performance overhead.
PROJECT
Research Project
Establishing foundational technologies for automated generation of parallel confidential computing code using generative AI.
INFORMATION
Project Information
- Research Area
- Creation of System Software for Society 5.0 by Integrating Fundamental Theories and System Platform Technologies
- Project Title
- AI-Based Automatic Code Generation for Parallel Confidential Computing
- Principal Investigator
- Shinobu Miwa (Associate Professor, Graduate School of Informatics and Engineering, The University of Electro-Communications)
- Project Period
- April 2026–March 2028
BACKGROUND
Background
Generative AI has made major advances in sequential code development, and its application to parallel code generation is accelerating. However, parallel confidential computing code running across multiple TEEs requires TEE-specific APIs and secure inter-TEE communication, so conventional parallel code cannot simply be used as-is.
CHALLENGES
Three Technical Challenges
We develop technologies for automating parallel confidential computing from three perspectives: creating code, verifying it, and securing it.
Advanced Testing Techniques
We enable comprehensive test generation and quality improvement for bug patterns specific to parallel execution, including races and deadlocks.
Security Verification
We evaluate AI robustness against attacks and vulnerabilities in generated code, feeding the security evaluation results back into code-generation AI.
RESEARCH ITEMS
Research Topics A–G
In FY2026, we begin with sequential confidential computing code and extend the resulting technologies to parallel confidential computing code in FY2027.
Evaluation of General-Purpose Generative AI
We task general-purpose generative AI systems such as ChatGPT with generating confidential computing code and analyze the results in terms of functional correctness, performance, and security.
Specialized Code-Generation AI
We retrain pretrained models to develop specialized AI for confidential computing code generation using code datasets and API specifications.
Code-Optimization AI
We develop AI agents that generate code satisfying required security specifications while reducing execution time.
Comprehensive Test Generation
We systematize known bug patterns and use agentic coding to automatically generate corresponding test cases.
Test Quality Improvement
We analyze and evaluate test code generated within the code-generation process to improve program-generation accuracy.
Theoretical Study of Confidential Computing
We analyze TEE security requirements and attacker capabilities from a computational complexity perspective to clarify the minimum guarantees required for confidential computing.
Security Evaluation of AI and Confidential Computing Code
We evaluate the security of AI and generated code using adversarial examples, backdoor attacks, static analysis, ProVerif, and related techniques.
SCHEME
Research Approach
The Miwa Group develops code-generation AI and shares generated code across the team. The Miwa Group evaluates performance, the Kashiwa Group verifies functionality, and the Yanai Group evaluates security. Results from each evaluation are fed back into AI development.
