Hanyang University team develops world's first Gaussian probability transistor to detect deepfake audio

A research team at Hanyang University has developed the foundational technology for a system capable of detecting deepfake audio.

Hanyang University announced Monday that a research team led by professor Yu Ho-cheon of the Department of Electronic Engineering, working with domestic and international collaborators, has developed the core technology behind a system that distinguishes deepfake audio from genuine speech in real time.

The team developed what it says is the world's first "electrical Gaussian transistor" for probability-based AI computing. The achievement involves directly implementing and freely adjusting an ideal Gaussian probability distribution at the hardware level using only a single semiconductor and a split-gate structure. The technology forms the basis for real-time discrimination between deepfake and genuine audio.

Conventional probability-based AI models have required repeated digital computations on CPUs or GPUs to calculate Gaussian distributions, consuming large amounts of power in the process.

In response, the research team developed a Single-Channel Gaussian-Mirroring Transistor (SC-GMT), applying a single semiconductor and split-gate structure.

The device achieves symmetric Gaussian characteristics through purely electrical control, and independently adjusts the distribution's amplitude (A), mean (μ) and standard deviation (σ).

The team then scaled the device into a printed circuit board (PCB)-based hardware system to build a Gaussian Naive Bayes (GNB) classifier capable of performing real-time probability computations. Using this classifier, the researchers implemented a system that distinguishes deepfake audio from genuine speech in real time. Experiments using the ASVspoof2019 dataset — an international benchmark challenge — confirmed high classification accuracy.

The achievement demonstrates the potential to perform probability distribution generation and computation directly within a single semiconductor device, tasks previously handled by software and digital circuits. The technology is expected to find broad application in next-generation AI semiconductor fields, including low-power edge AI, probability-based AI and deepfake detection.

"This research presents a new semiconductor platform capable of directly generating and computing probability distributions at the device level," Yu said. "We expect it to be widely applied across a range of next-generation AI semiconductor fields, including low-power edge AI, probability-based AI and deepfake detection."