About this Abstract |
| Meeting |
2027 TMS Annual Meeting & Exhibition
|
| Symposium
|
Advanced Biomaterials and Implants
|
| Presentation Title |
Light-Induced True Randomness for Secure and Data-Augmented Medical AI |
| Author(s) |
Juhyung Seo, Hocheon Yoo, Taehyun Park |
| On-Site Speaker (Planned) |
Juhyung Seo |
| Abstract Scope |
Medical artificial intelligence depends on digital images but faces two major challenges, image manipulation and limited training datasets constrained by patient privacy. Here, we present a dual-function framework using light-induced true random number generators for medical image authentication and synthetic data generation. A CuV2O6/SnO2 photospike true random number generator produces high-entropy ternary outputs through stochastic charge trapping and detrapping, showing near-ideal uniformity and inter-Hamming distance and passing all NIST randomness tests. The generated codes are embedded as visually imperceptible hidden layers in medical images, enabling pixel-level detection of unauthorized or AI-based modifications. In parallel, an arc-discharge light-induced true random number generator driven by competing positive and negative photocurrents serves as a physical randomness source for StyleGAN2-ADA to generate synthetic pneumothorax chest X-ray images. Compared with pseudo-random inputs, the true-randomness-driven model achieves lower FID and KID values, demonstrating improved image quality and supporting secure, privacy-conscious medical AI with expanded training datasets. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Electronic Materials, Nanotechnology, Thin Films and Interfaces |