The DeepFake, which generates synthetic content, has sparked a revolution in the fight against deception and forgery. However, most existing DeepFake detection methods mainly focus on improving detection performance with high-quality data while ignoring low-quality synthetic content that suffers from high compression. To address this issue, we propose a novel High-Frequency Enhancement framework, which leverages a learnable adaptive high-frequency enhancement network to enrich weak high-frequency information in compressed content without uncompressed data supervision. The framework consists of three branches, i.e., the Basic branch with RGB domain, the Local High-Frequency Enhancement branch with Block-wise Discrete Cosine Transform, and the Global High-Frequency Enhancement branch with Multi-level Discrete Wavelet Transform. Among them, the local branch utilizes the Discrete Cosine Transform coefficient and channel attention mechanism to indirectly achieve adaptive frequency-aware multi-spatial attention, while the global branch supplements the high-frequency information by extracting coarse-to-fine multi-scale high-frequency cues and cascade-residual-based multi-level fusion by Discrete Wavelet Transform coefficients. In addition, we design a Two-Stage Cross-Fusion module to effectively integrate all information, thereby greatly enhancing weak high-frequency information in low-quality data. Experimental results on FaceForensics++, Celeb-DF, and OpenForensics datasets show that the proposed method outperforms the existing state-of-the-art methods and can effectively improve the detection performance of DeepFakes, especially on low-quality data. The code is available here.

DeepFake detection based on high-frequency enhancement network for highly compressed content

Marcialis, Gian Luca
Supervision
;
2024-01-01

Abstract

The DeepFake, which generates synthetic content, has sparked a revolution in the fight against deception and forgery. However, most existing DeepFake detection methods mainly focus on improving detection performance with high-quality data while ignoring low-quality synthetic content that suffers from high compression. To address this issue, we propose a novel High-Frequency Enhancement framework, which leverages a learnable adaptive high-frequency enhancement network to enrich weak high-frequency information in compressed content without uncompressed data supervision. The framework consists of three branches, i.e., the Basic branch with RGB domain, the Local High-Frequency Enhancement branch with Block-wise Discrete Cosine Transform, and the Global High-Frequency Enhancement branch with Multi-level Discrete Wavelet Transform. Among them, the local branch utilizes the Discrete Cosine Transform coefficient and channel attention mechanism to indirectly achieve adaptive frequency-aware multi-spatial attention, while the global branch supplements the high-frequency information by extracting coarse-to-fine multi-scale high-frequency cues and cascade-residual-based multi-level fusion by Discrete Wavelet Transform coefficients. In addition, we design a Two-Stage Cross-Fusion module to effectively integrate all information, thereby greatly enhancing weak high-frequency information in low-quality data. Experimental results on FaceForensics++, Celeb-DF, and OpenForensics datasets show that the proposed method outperforms the existing state-of-the-art methods and can effectively improve the detection performance of DeepFakes, especially on low-quality data. The code is available here.
2024
DeepFake detection; Discrete cosine transform; Discrete wavelet transform; High-frequency enhancement; Low-quality deepFake
File in questo prodotto:
File Dimensione Formato  
J41_ESWA_2024_PP.pdf

Solo gestori archivio

Descrizione: articolo online
Tipologia: versione editoriale (VoR)
Dimensione 8.24 MB
Formato Adobe PDF
8.24 MB Adobe PDF   Visualizza/Apri   Richiedi una copia
ESWA-preprint-JieGao_cover.pdf

accesso aperto

Descrizione: Pre-print
Tipologia: versione pre-print
Dimensione 5.33 MB
Formato Adobe PDF
5.33 MB Adobe PDF Visualizza/Apri

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11584/394503
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 5
  • ???jsp.display-item.citation.isi??? ND
social impact