Fooling State-of-the-Art Deepfake Detection with High-Quality Deepfakes

View on arXiv ← Back to list

Authors: Arian Beckmann, Anna Hilsmann, Peter Eisert

Published: 2023-05-09 09:08:49+00:00

AI Summary

This paper investigates the limitations of deepfake detectors trained solely on existing datasets. The authors generate high-quality deepfakes using a novel autoencoder and blending technique, demonstrating that a state-of-the-art detector performs poorly on these realistic fakes. Fine-tuning the detector on these high-quality fakes improves performance, highlighting the need for such data in training robust detectors.

Abstract

Due to the rising threat of deepfakes to security and privacy, it is most important to develop robust and reliable detectors. In this paper, we examine the need for high-quality samples in the training datasets of such detectors. Accordingly, we show that deepfake detectors proven to generalize well on multiple research datasets still struggle in real-world scenarios with well-crafted fakes. First, we propose a novel autoencoder for face swapping alongside an advanced face blending technique, which we utilize to generate 90 high-quality deepfakes. Second, we feed those fakes to a state-of-the-art detector, causing its performance to decrease drastically. Moreover, we fine-tune the detector on our fakes and demonstrate that they contain useful clues for the detection of manipulations. Overall, our results provide insights into the generalization of deepfake detectors and suggest that their training datasets should be complemented by high-quality fakes since training on mere research data is insufficient.

Key findings

A state-of-the-art deepfake detector showed significantly reduced accuracy (from 97.4% to 26.7%) when tested on the high-quality deepfakes. Fine-tuning the detector on these fakes improved its performance, indicating that including high-quality fakes in training datasets is crucial for robust deepfake detection. The fine-tuned models still performed well on existing benchmarks (FFPP and CDF).

Approach

The authors propose a novel autoencoder architecture for face swapping, incorporating an improved face blending technique. They generate 90 high-quality deepfakes using this approach and evaluate their impact on a state-of-the-art deepfake detector, followed by fine-tuning the detector with the generated data.

Datasets

Actors subset of a deepfake detection dataset (containing videos of 28 actors); FaceForensics++ (FFPP); Celeb-DF v2 (CDF)

Model(s)

A novel dual-decoder autoencoder using EfficientNet-B4 as the encoder; RealForensics detector

Author countries

Germany

← Previous