We proposed a general-purpose deepfake image detector called Time Step Generating (TSG). Unlike existing methods, TSG is independent of pretraining model reconstruction abilities, specific datasets, or sampling algorithms. By feeding images to a pre-trained diffusion model at a fixed timestep, TSG captures the differences in noise-prediction behavior between real and generated images within a single forward pass.
To remove the reliance on a manually chosen timestep, TSG++ consolidates multi-timestep diffusion features through lightweight fine-tuning, improving robustness and generalization without additional inference cost.
We further extended this deepfake detection technique to fake video detection, enabling localization of manipulated frames and forged regions.
Published: ACM MMAsia 2025 (Oral)
Ziyue Zeng
Ph.D. Student
Pursuing socially meaningful, grounded research with enduring passion and diligence.