Paper: Proactive Defense Benchmark against Deepfake Generation
Authors: Joonhyuk Baek, Wonjune Seo, Jae-yun Kim, Saerom Park, Hoki Kim
Venue: ICML 2026
Our lab paper “Proactive Defense Benchmark against Deepfake Generation” has been accepted to ICML 2026, one of the most prestigious venues in machine learning.
As deepfake generation continues to advance rapidly, there is growing consensus that detecting fakes after the fact is no longer sufficient. This has driven research into proactive defenses that perturb source images in advance to make deepfake generation difficult — but the absence of a unified evaluation protocol has made fair comparison between methods nearly impossible.
This work establishes a comprehensive benchmark that evaluates proactive deepfake defenses under a unified protocol. The analysis quantitatively reveals a critical trade-off between fidelity preservation and identity protection metrics, and uses these findings to chart a direction for next-generation defenses with stronger generalization.
The paper will be presented at ICML 2026.