Abstract:Splicing forgery, which manipulates images by copying regions from donor images and pasting them to host images, is one of the common types of image forgery in life, where the copied regions include object regions or background regions. In order to accurately detect these forgery regions, the most mainstream approach is to use an encoder-decoder network architecture that extracts enough manipulation traces to determine whether each pixel of the input image has been spliced or not. However, due to the limited receptive field of such networks, only local manipulation traces can be learned, and therefore some large object area forgery and background forgery cannot be well localized. To address these issues, in this paper, an end-to-end splicing detection framework is proposed, which includes localization network L-Net, manipulation traces attention network MTA-Net, and adaptive multi-scale fusion module. The localization network L-Net is designed as an encoder-decoder network to extract local manipulation traces for each pixel and implement localization of splicing areas. MTA-Net uses the proposed content-remove convolutional layer (CRCL) to suppress image content information that would hinder the network from learning to manipulate traces, and then uses subsequent convolutional layers to extract features to discriminate whether the input image is a spliced image or not. In this process, the regions in the feature map of the convolutional layers with large activation values are the ones that contain global manipulation traces. These global manipulation traces are fused with the local manipulation traces learned by L-Net through the proposed adaptive multi-scale fusion module (AMSFM), thus allowing L-Net to effectively handle object forgery and background region forgery images of various sizes. Ablation experiments showed an increase of 4.6% and 3.9% in F1-score and MCC after the introduction of MTA-Net and AMSFM, respectively The splicing region detection performance on three standard datasets, CASIA, COLUMB, and CARVALHO, shows that the proposed method outperforms the state-of-the-art methods for both object forgery and background forgery, and is more robust to post-processing methods such as JPEG compression and noise addition.

Adversarial Attack on Deep Learning-Based Splice Localization

Different-quality Re-demosaicing in Digital Image Forensics

Image splicing localization based on re-demosaicing

End-to-end Image Splicing Localization Based on Multi-Scale Features and Residual Refinement Module

Frequency-constrained transferable adversarial attack on image manipulation detection and localization

Feature enhancement and supervised contrastive learning for image splicing forgery detection

Joint Manipulation Trace Attention Network and Adaptive Fusion Mechanism for Image Splicing Forgery Localization

Analysis of adversarial attacks against CNN-based image forgery detectors

Adversarial mimicry attacks against image splicing forensics: An approach for jointly hiding manipulations and creating false detections

Adversarial Learning for Constrained Image Splicing Detection and Localization Based on Atrous Convolution

Video Splicing Detection and Localization Based on Multi-Level Deep Feature Fusion and Reinforcement Learning

Toward effective image forensics via a novel computationally efficient framework and a new image splice dataset

Adversarial Analysis for Source Camera Identification

Towards Effective Image Forensics via A Novel Computationally Efficient Framework and A New Image Splice Dataset

Transferable Adversarial Attack on Image Tampering Localization

On the Transferability of Adversarial Examples Against CNN-Based Image Forensics

Detecting Localized Adversarial Examples: A Generic Approach Using Critical Region Analysis

Evading Detection Actively: Toward Anti-Forensics against Forgery Localization

Fighting Fake News: Image Splice Detection via Learned Self-Consistency

Mitigating Adversarial Perturbations Via Weakly Supervised Object Location and Regions Recombination

RegionSparse: Leveraging Sparse Coding and Object Localization to Counter Adversarial Attacks