Adversarial Perturbations Fool Deepfake Detectors

Authors: Apurva Gandhi, Shomik Jain

Published: 2020-03-24 00:54:02+00:00

Comment: To appear in the proceedings of the International Joint Conference on Neural Networks (IJCNN 2020)

AI Summary

This paper investigates the vulnerability of deepfake image detectors to adversarial perturbations and proposes two defense mechanisms. It demonstrates that common deepfake detectors, which typically achieve high accuracy on unperturbed images, can be severely degraded to less than 27% accuracy when confronted with adversarially perturbed deepfakes. The study evaluates Lipschitz regularization and Deep Image Prior (DIP) as countermeasures, with DIP showing significant promise in restoring detection accuracy.

Abstract

This work uses adversarial perturbations to enhance deepfake images and fool common deepfake detectors. We created adversarial perturbations using the Fast Gradient Sign Method and the Carlini and Wagner L2 norm attack in both blackbox and whitebox settings. Detectors achieved over 95% accuracy on unperturbed deepfakes, but less than 27% accuracy on perturbed deepfakes. We also explore two improvements to deepfake detectors: (i) Lipschitz regularization, and (ii) Deep Image Prior (DIP). Lipschitz regularization constrains the gradient of the detector with respect to the input in order to increase robustness to input perturbations. The DIP defense removes perturbations using generative convolutional neural networks in an unsupervised manner. Regularization improved the detection of perturbed deepfakes on average, including a 10% accuracy boost in the blackbox case. The DIP defense achieved 95% accuracy on perturbed deepfakes that fooled the original detector, while retaining 98% accuracy in other cases on a 100 image subsample.

Key findings

Deepfake detectors (VGG and ResNet) achieved over 95% accuracy on unperturbed images but dropped to less than 27% on adversarially perturbed deepfakes. Lipschitz regularization provided a modest improvement in robustness, increasing detection accuracy by approximately 10% in the blackbox case. The Deep Image Prior (DIP) defense was highly effective, achieving 95% accuracy on perturbed deepfakes that initially fooled detectors, while maintaining 98% accuracy on other image types, albeit with significant computational overhead.

Approach

The authors first create adversarial perturbations for deepfake images using FGSM and Carlini and Wagner L2 norm attacks in both blackbox and whitebox settings to evaluate their effectiveness against VGG and ResNet detectors. They then propose two defense strategies: Lipschitz regularization, which constrains the detector's gradient to increase robustness, and Deep Image Prior (DIP), an unsupervised method using a generative CNN to remove perturbations from images before classification.

Datasets

CelebA (for real images), and a dataset of 5,000 fake images generated using Few-Shot Face Translation GAN.

Model(s)

UNKNOWN

Author countries

USA

← Previous