From Creation to Detection: How Dataset Composition and Simple Augmentation Influence Deepfake Training

Abstract

Deepfakes are sophisticated, AI-generated alterations of images and videos that pose significant threats to cybersecurity, particularly with face-swapping techniques that can deceive and spread misinformation. A major limitation in current deepfake detection strategies is that the deepfakes used to train these models are often of lower quality than those encountered in real-world scenarios [1]. This weakens model performance when tested against more sophisticated media alterations.

To bridge this gap, deepfake detection datasets must evolve to include high‑quality deepfakes that better reflect real‑world threats. This study examines popular datasets such as Celeb‑DF and DF‑1.0, which revealed that despite efforts toward attribute variability, these datasets often lack demographic and facial diversity. Our study uses FaceSwap to explore how dataset composition and simple augmentation (horizontal mirroring) relate to training behavior and output quality in a small set of faceswaps. We observed that augmented datasets were associated with slightly lower faceloss—a training‑stage metric that reflects how well the model reconstructs or blends facial regions—and noted attribute‑dependent differences across the subjects tested. These findings are preliminary and reflect an exploratory analysis across a small set of subjects and configurations. Because faceloss functions as a training‑stage proxy rather than a perceptual or detection metric, broader replication with larger and more diverse datasets would strengthen the statistical power of future evaluations. Nevertheless, the patterns observed in this work highlight promising directions for improving dataset balance and transparency and suggest practical strategies for mitigating attribute‑driven skews in both deepfake creation and detection.

The study further investigates how dataset composition affects deepfake generation by evaluating two factors: (1) the impact of horizontal‑mirroring augmentation, which aims to increase facial‑orientation variability, and (2) how FaceSwap performs on subjects with different attributes to reveal potential skews in the deepfake generation process.