notesum.ai

Published at November 12

Adaptive Meta-Learning for Robust Deepfake Detection: A Multi-Agent Framework to Data Drift and Model Generalization

cs.AI

Released Date: November 12, 2024

Authors: Dinesh Srivasthav P¹, Badri Narayan Subudhi¹

Aff.: ¹Indian Institute of Technology Jammu

Arxiv: http://arxiv.org/abs/2411.08148v1

Refer to caption

Dataset	Total Real	Total Fake	Class Count	Deepfake Type	Fake Samples Generation Method	Real Samples Collection Method
DeepFakeFace [4]	30k	90k	Real - 1 Fake - 3 Total - 4	Synthetic images of celebrities	• Stable Diffusion v1.5 (30k) • InsightFace toolbox (30k) • Stable Diffusion Inpainting (30k)	IMDB-Wiki dataset (30k)
DGM [52]	81.5k	125.8k	Real - 4 Fake - 4 Total - 8	Face swap manipulations, face attribute manipulations	• HFGI (28.5k) • Infoswap (44k) • Simswap (22.5k) • StyleClip (30.8k)	• BBC (15k) • The Guardian (43k) • USA Today (13.9k) • The Washington Post (9.6k)
iFakeFaceDB [53]	-	87k (63k+24k)	Fake - 2 Total - 2	Removing GAN fingerprints from synthetic images	Generated by StyleGAN, and transformed with GANprintR, applied on following DBs: • 100k-faces db (24k) • TPDNE db (62k)	-
CocoGlide⁵⁵5https://github.com/grip-unina/TruFor [54]	513	513	Real - 1 Fake - 1 Total - 2	Image splicing; Diffusion based	GLIDE (text-guided diffusion model)	COCO dataset
DF40 [55]	4k	326.5k	Real - 3 Fake - 14 Total - 17	Face swapping, Full face synthesis, Face edit	• MidJourney (1k) • WhichFaceIsReal (1k) • VQGAN (31.8k) • StyleGAN2 (31.8k) • StyleGAN3 (31.8k) • StyleGANXL (31.8k) • StyleCLIP (31k) • StarGanV2 (2k) • SiT (31.8k) • DiT (31.5k) • RDDM (17.6k) • DDIM (31.8k) • • e4e (50.6k) • CollabDiff (1k)	• StarGANv2 (2k) • WhichFaceIsReal (1k) • StyleCLIP (1k)
OpenForensics-based⁶⁶6https://www.kaggle.com/datasets/manjilkarki/deepfake-and-real-images/data [56]	95k	95k	Real - 1 Fake - 1 Total - 2	Multi-face forgery; Forged face image synthesis; Face swapping; Various Perturbations and augmentations	GAN; Image transformation; Poisson blending; Color adaptation	Google Open Images