ML Security PapersScaling Laws for Deepfake Detection
Wenhao Wang 1, Longqi Cai 2, Taihong Xiao 2, Yuxiao Wang 2, Ming-Hsuan Yang 2
Published on arXiv
2510.16320
Output Integrity Attack
OWASP ML Top 10 — ML09
Key Finding
Detection error follows a predictable power-law decay with respect to the number of real domains or deepfake methods, with no saturation observed across 51 domains and 102 generation methods, enabling forecasting of data requirements to hit target detection performance
ScaleDF
Novel technique introduced
This paper presents a systematic study of scaling laws for the deepfake detection task. Specifically, we analyze the model performance against the number of real image domains, deepfake generation methods, and training images. Since no existing dataset meets the scale requirements for this research, we construct ScaleDF, the largest dataset to date in this field, which contains over 5.8 million real images from 51 different datasets (domains) and more than 8.8 million fake images generated by 102 deepfake methods. Using ScaleDF, we observe power-law scaling similar to that shown in large language models (LLMs). Specifically, the average detection error follows a predictable power-law decay as either the number of real domains or the number of deepfake methods increases. This key observation not only allows us to forecast the number of additional real domains or deepfake methods required to reach a target performance, but also inspires us to counter the evolving deepfake technology in a data-centric manner. Beyond this, we examine the role of pre-training and data augmentations in deepfake detection under scaling, as well as the limitations of scaling itself.
Key Contributions
- ScaleDF: the largest deepfake detection dataset with 5.8M+ real images from 51 domains and 8.8M+ fake images from 102 deepfake generation methods across 5 forgery categories
- Discovery of power-law scaling laws for deepfake detection: detection error follows 1 − AUC = A·N^(−α) as the number of real domains or deepfake methods increases, with no observed saturation
- Empirical analysis of how pre-training and data augmentation interact with data scale, and identification of limitations where performance saturates beyond ~10M images at fixed domain/method counts
🛡️ Threat Analysis
Deepfake detection (AI-generated face content detection) is explicitly listed under ML09. The paper's primary contributions — ScaleDF dataset and scaling law analysis — directly advance the capability to detect AI-generated/manipulated imagery, which is the core output integrity concern of ML09.