Limits of Deepfake Detection: A Robust Estimation Viewpoint

View on arXiv ← Back to list

Authors: Sakshi Agarwal, Lav R. Varshney

Published: 2019-05-09 09:01:08+00:00

AI Summary

This paper formulates deepfake detection as a hypothesis testing problem. It uses robust statistics to bound the error probability of various GAN implementations and simplifies these bounds using a Euclidean approximation for low error regimes. Finally, it establishes relationships between error probability and epidemic thresholds in networks.

Abstract

Deepfake detection is formulated as a hypothesis testing problem to classify an image as genuine or GAN-generated. A robust statistics view of GANs is considered to bound the error probability for various GAN implementations in terms of their performance. The bounds are further simplified using a Euclidean approximation for the low error regime. Lastly, relationships between error probability and epidemic thresholds for spreading processes in networks are established.

Key findings

The error probability of deepfake detection decreases exponentially with image resolution and inversely with the accuracy of the GAN used. Bounds derived using different distance metrics show that for higher GAN accuracy (lower OPT), detection becomes exponentially harder in the Neyman-Pearson framework and polynomially harder in the Bayesian framework.

Approach

The authors frame deepfake detection as a hypothesis testing problem, leveraging robust statistics to bound the error probability of GAN-generated images. They analyze different distance metrics (KL divergence, total variation distance, etc.) to derive error bounds, further simplified using Euclidean approximations for low-error scenarios.

Datasets

UNKNOWN

Model(s)

UNKNOWN

Author countries

USA

← Previous