Jeonnam National University Develops Preemptive Deepfake Defense Technology "DeepProtect"

Jeonnam National University campus.

Jeonnam National University announced on March 5 that Professor Yoo Seokbong’s research team has developed a technology called “DeepProtect,” which protects users’ faces in advance, making it difficult to generate deepfakes.

Recently, the rapid advancement of “deepfake (face-swapping)” technology—using artificial intelligence (AI) to naturally composite another person’s face in movies and games—has led to an increase in serious crimes such as unauthorized face theft, digital sex crimes, fraud, and the creation of fake videos.

While existing countermeasures have mostly relied on “post-detection” methods to identify deepfakes after they are created, the proposed technology distinguishes itself by providing preemptive defense. It processes photos for protection before they are uploaded to social media, so that even if someone tries to use the photo to swap faces, the resulting deepfake will not appear natural.

The research team implemented a dual strategy that simultaneously defends both the entire face (global) and specific regions (local). The proposed “identity blending” technique does not completely remove unique facial features; instead, it subtly mixes features from several similar faces—identified through a filter bank-based search—thereby blurring the overall identity information of the face.

Additionally, the team introduced an “attribute distortion” technique. If a user specifies certain facial regions—such as eyes, nose, or mouth—through text prompts, the system identifies the direction of identity in the AI’s internal representation space related to those regions. It then inserts a subtle watermark signal in that direction, keeping the original image natural while ensuring that, during deepfake generation, those regions appear unnatural.

Experimental results showed that images processed with DeepProtect maintained high naturalness in both quantitative and human qualitative evaluations, and demonstrated a high defense success rate against various state-of-the-art facial deepfake models.

This research was jointly authored by Seunghyuk Baek and Eunki Lee, both master’s students at the Visual Intelligence Media Lab in the Department of AI Convergence at Jeonnam National University. Professors Yoo Seokbong and Kim Hyungil jointly led the research. The study has been accepted for presentation at CVPR 2026, the world’s leading academic conference in AI computer vision, and will be officially presented in June this year.

Co-first author Seunghyuk Baek, in addition to this paper, is growing as a globally competitive researcher with achievements such as a paper accepted at ICRA 2026 in the field of AI adversarial robustness. Meanwhile, researcher Eunki Lee, based on outstanding research achievements during the master’s program, was recently appointed as a new researcher at the Honam Regional Research Center of the Electronics and Telecommunications Research Institute (ETRI) in Korea.

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