Alapfogalmak
Deep image restoration techniques can effectively mitigate the quality degradation caused by simple anti-forensics methods, making it harder for state-of-the-art image forgery detection models to identify manipulated images.
Kivonat
This study explores the use of deep image restoration techniques to enhance image anti-forensics. The authors first identify common anti-forensics methods such as blurring, resizing, JPEG compression, and noise addition, which are often used for data augmentation in training image forgery detection models. However, these simple methods not only make it difficult to detect image manipulations, but also degrade the overall image quality.
To address this issue, the authors propose a two-step approach that combines anti-forensics methods with deep image restoration techniques. Specifically, they evaluate the following methods:
Blur&Sharp: Applying Gaussian blurring followed by sharpening to restore image quality.
Downsize&Upsize: Asymmetrically downscaling the image and then upscaling it back to the original size.
JPEG Compression&JPEG Compression Artifact Removal: Applying JPEG compression and then using a deep learning model (FBCNN) to remove the compression artifacts.
Gaussian Noise&Denoise: Adding Gaussian noise to the image and then using a deep learning model (Restormer) to remove the noise.
Downscale&Upscale: Downscaling the image by half and then using a deep learning super-resolution model (SwinFIR) to upscale it back to the original size.
The authors evaluate the impact of these methods on image quality using various metrics (PSNR, SSIM, BRISQUE) and test their effectiveness against two state-of-the-art image forgery detection models (Trufor and Early Fusion). The results show that the proposed methods can significantly reduce the accuracy and recall of these models, making it harder for them to detect manipulated images.
The authors conclude that the inclusion of advanced anti-forensics methods, such as those presented in this study, in the data creation and model training processes is necessary to improve the robustness of image forgery detection models against real-world image manipulations.
Statisztikák
The image quality metrics for the different anti-forensics methods on the COVERAGE and DSO-1 datasets are as follows:
COVERAGE:
Raw Images: PSNR - -, SSIM - -, BRISQUE - 18.96
Blur&Sharp: PSNR - 32.69, SSIM - 0.940, BRISQUE - 37.15
Downsize&Upsize: PSNR - 32.08, SSIM - 0.944, BRISQUE - 34.26
Gaussian Noise&Denoise (Sigma=15): PSNR - 37.35, SSIM - 0.961, BRISQUE - 36.04
Gaussian Noise&Denoise (Sigma=25): PSNR - 34.91, SSIM - 0.942, BRISQUE - 39.20
JPEG Compression&JPEG CAR (QF=50): PSNR - 35.68, SSIM - 0.955, BRISQUE - 38.87
JPEG Compression&JPEG CAR (QF=70): PSNR - 37.30, SSIM - 0.966, BRISQUE - 35.71
DSO-1:
Raw Images: PSNR - -, SSIM - -, BRISQUE - 13.58
Blur&Sharp: PSNR - 38.92, SSIM - 0.951, BRISQUE - 24.10
Downsize&Upsize: PSNR - 39.70, SSIM - 0.964, BRISQUE - 30.97
Gaussian Noise&Denoise (Sigma=15): PSNR - 37.72, SSIM - 0.926, BRISQUE - 35.92
Gaussian Noise&Denoise (Sigma=25): PSNR - 36.03, SSIM - 0.906, BRISQUE - 43.67
JPEG Compression&JPEG CAR (QF=50): PSNR - 37.29, SSIM - 0.925, BRISQUE - 41.69
JPEG Compression&JPEG CAR (QF=70): PSNR - 38.65, SSIM - 0.940, BRISQUE - 33.29
Downscale&Upscale: PSNR - 38.61, SSIM - 0.956, BRISQUE - 30.70