Sulthan Abiyyu Hakim et al., supervised by Novanto Yudistira. Jurnal Teknologi Informasi dan Ilmu Komputer, Vol. 11 No. 3, June 2024.

Why it matters

AI-generated imagery is useful in research and industry, and corrosive everywhere that an image functions as evidence: legal proceedings, political reporting, journalism. Being able to ask whether an image is authentic is becoming infrastructure.

Approach

We fine-tuned residual networks, and the research focus was on the fine-tuning technique rather than the architecture:

  • A warm-up-based learning rate scheduler followed by a linear scheduler
  • Gradient accumulation
  • Image augmentation

Results

ResNet-152 performed best: F1 score 0.963, loss 0.08.

The finding worth keeping is that careful fine-tuning technique on an established architecture outperformed reaching for something more exotic. The scheduler and augmentation choices moved the number more than the backbone did.