KTC Takes Part in Developing the World's First Physically Resilient Deep Learning Accelerator
KTC Takes Part in Developing the World's First Physically Resilient Deep Learning Accelerator
Korea Testing Certification (KTC) is advancing the national information protection infrastructure by participating in the project titled "Vulnerability Verification and Response Technology Development for On-Device Deep Learning Accelerator Design based on Physical Channels," organized by Institute for Information & Communications Technology Planning & Evaluation(IITP).
In collaboration with six institutions, including the leading institution Korea University, as well as Kookmin University, Pusan National University, Sookmyung Women's University, TS&Labs Inc., and SecureBox Corp., KTC is jointly executing the project.
* Deep Learning: A machine learning technique inspired by the human brain that enables systems to learn from data.
The goal of this project is to develop the world's first board (chip) capable of accurate operations and deep learning in the face of malicious physical attacks on deep learning accelerators (such as graphics processing units, programmable semiconductors, and custom semiconductors) embedded in devices.
Deep learning accelerators find widespread use in various industries, including smartphones, wearable devices, smart factories, smart grids, smart homes, and autonomous driving. Consequently, enhancing the security of deep learning devices has become increasingly important, as a breach of crucial information in these devices could lead to societal damages, such as personal information leakage and replication.
Through this project, technology is being developed to make deep learning equipment resistant to physical channel-based attacks, thereby minimizing the threats of personal information and deep learning model data leakage.
Physical attacks on deep learning accelerators include fault injection attacks that induce errors by injecting stimuli into the board using electromagnetic waves or lasers during operation, and side-channel attacks that analyze and exploit physical information such as power consumption and electromagnetic waves to capture encryption keys. Side-channel attacks are employed for illegal activities, such as extracting and replicating confidential information from financial IC cards and transportation cards.
Since its initiation in 2021, the project has made significant progress in its third year. It has successfully developed a low-power, small-size board that checks vulnerabilities in deep learning algorithms while minimizing resource usage. Additionally, it has surpassed its project goals, publishing five papers in scientific journals, including those indexed in the Science Citation Index (SCI). Building on this progress, one domestic patent has been registered, and two international patents, along with three additional domestic patents, have been applied for.
In the project's final year in 2024, KTC will finalize the development of safety verification methods for deep learning accelerator physical channels. This involves conducting vulnerability simulation and verification of boards resistant to physical attacks, providing evidence of their safety.
KTC President Sungil Ahn remarked, "Deep learning services are widely utilized in various industries, highlighting the increasing importance of enhancing security," and added, "Through this project, we aim to contribute to leading the Korean market in deep learning accelerator security technology."
