A New Benchmark and Model for Challenging Image Manipulation Detection

TL;DR

The paper introduces CIMD, a challenging image manipulation detection benchmark with two subsets for editing-based and compression-based tampering, designed to test detection of small-forgery regions and double compression with identical Quality Factors. It also proposes a two-branch IMD model that fuses an RGB anomaly-detection pathway with a frequency-domain compression-artifact learner, leveraging HRNet backbones, ASPP, and interactive attention, with adaptive heatmap fusion. Experimental results on CIMD show the proposed method significantly outperforms existing SoTA IMD approaches on both CIMD-R and CIMD-C, including ablations that demonstrate the benefit of combining both branches. The work provides a rigorous, high-quality dataset and a robust detection framework with practical forensic relevance for detecting subtle edits and same-QF double-compression artifacts.

Abstract

The ability to detect manipulation in multimedia data is vital in digital forensics. Existing Image Manipulation Detection (IMD) methods are mainly based on detecting anomalous features arisen from image editing or double compression artifacts. All existing IMD techniques encounter challenges when it comes to detecting small tampered regions from a large image. Moreover, compression-based IMD approaches face difficulties in cases of double compression of identical quality factors. To investigate the State-of-The-Art (SoTA) IMD methods in those challenging conditions, we introduce a new Challenging Image Manipulation Detection (CIMD) benchmark dataset, which consists of two subsets, for evaluating editing-based and compression-based IMD methods, respectively. The dataset images were manually taken and tampered with high-quality annotations. In addition, we propose a new two-branch network model based on HRNet that can better detect both the image-editing and compression artifacts in those challenging conditions. Extensive experiments on the CIMD benchmark show that our model significantly outperforms SoTA IMD methods on CIMD.

Figures (6)

• Figure 1: Sample images of our dataset and comparison of image manipulation detection results with recent mainstream methods. The first three rows show manipulation of region copy-move, splicing and removal, respectively. The last row shows double-compressed splicing with the same Quality Factor (QF). Our method achieves the new state-of-the-art in detecting challenging manipulation cases.
• Figure 2: Overview of the proposed two-branch architecture. RGB stream can detect anomalous features, while frequency stream is able to learn compression artifacts by feeding the image to the compression artifacts learning model, as depicted in Fig. \ref{['fig:JPEG']}. The ASPP in Fig. \ref{['fig:atten']}(a) is appended to each of the outputs, and channel attention and spatial attention in Fig. \ref{['fig:atten']}(b)(c) interactively perform between each scale output to improve the detection performance under small manipulation.
• Figure 3: DCT coefficient histograms from the (0,1) position generated from a raw image under different compression processes. The range of X-axis is [-20, 20].
• Figure 4: Visualization of DCT coefficients for each recompression for a repeatedly compressed image under QF 80. The number below shows recompression counts. Black pixels indicate unaltered DCT coefficients. White pixels indicate the unstable region where DCT coefficients change after compression, which gradually focus on the tampered region as the count increases.
• Figure 5: The compression artifact learning module. Three types ( de-quantized, quantized, and residual quantized) of DCT features are fed into the backbone to learn double compression artifacts in cases whether the QFs are the same or not.