'Development of AI Semiconductor for Deepfake Implementation on Mobile'

[Asia Economy Reporter Junho Hwang] Domestic researchers have developed a technology that enables the construction of Generative Adversarial Networks (GANs), which generate images by synthesizing or changing styles using artificial intelligence (AI), even on low-power devices such as smartphones or tablet PCs.

The Korea Advanced Institute of Science and Technology (KAIST) announced on the 6th that Professor Hoejun Yoo's research team from the Department of Electrical Engineering has developed GANPU, an AI semiconductor capable of processing GANs with low power consumption.

This semiconductor implements GANs, which were previously impossible to compute on mobile devices like smartphones. GAN is a next-generation deep learning technology where a deep learning network that generates fakes and another that discriminates them compete and learn from each other. Using this, AI can learn independently and generate images without human data input. Recently, it has gained attention as a technology capable of generating images that can even deceive humans, such as Deep Fake.

The research team developed GANPU to enable GANs with low power consumption through core technologies such as adaptive workload allocation, maximizing the utilization of input/output sparsity, and zero-pattern prediction using only the exponent part. As a result, GANPU achieves 4.8 times higher energy efficiency compared to existing semiconductors.

In particular, the team demonstrated an application technology where users can directly edit photos taken with a tablet camera using GANPU. When users input additions, deletions, or modifications for 17 features such as hair, glasses, and eyebrows on faces in the photo, GANPU generates images in real time according to the requests.

Professor Hoejun Yoo said, "This research is significant in that we developed an AI semiconductor capable of not only inference but also learning on a single chip, supporting multiple deep learning networks simultaneously. It greatly expands the scope of AI use on mobile devices and is expected to be applied diversely to GAN-related applications such as image style transfer, video synthesis, and image restoration."

The research results were recently presented at the International Solid-State Circuits Conference (ISSCC) held in San Francisco, USA.

This content was produced with the assistance of AI translation services.