Training Data Improvement for Image Forgery Detection using Comprint

Abstract

Manipulated images are a threat to consumers worldwide, when they are used to spread disinformation. Therefore, Comprint enables forgery detection by utilizing JPEGcompression fingerprints. This paper evaluates the impact of the training set on Comprint's performance. Most interestingly, we found that including images compressed with low quality factors during training does not have a significant effect on the accuracy, whereas incorporating recompression boosts the robustness. As such, consumers can use Comprint on their smartphones to verify the authenticity of images.

JADE OWL: JPEG 2000 Forensics by Wavelet Offset Consistency Analysis

“…ManTraNet [9], [45] -0.006 0.189 Noisesniffer [7] -0.009 0.035 Noiseprint [10] 0.006 0.080 Comprint [46] 0.005 0.070 Splicebuster [47] -0.005 0.205 AdaCFA [5] 0.178 0.324 ZERO [8] 0.000 0.000 As can be seen here, the proposed method can still work on 8-bits images, and even against a small JPEG compression, although its performances are expectedly diminished.…”

“…For the purpose of comparison, we measure our method's performance against several well-known, publicly available SOTA techniques. These include Splicebuster [47], Noiseprint [10], Noisesniffer [7], Comprint [46], ManTraNet [9], [45], AdaCFA [5] and ZERO [8], [43]. We quantify the results using Matthew's Correlation Coefficient (MCC), a metric that varies from -1 to 1; here, a score of 1 signifies a flawless detection, -1 its inverse, and 0 suggests an absence of correlation between the detection results and the ground truth.…”

“…In Table III, we analyse the results of the method using the finest scale only, the second-finest scale only, and both scales, depending on the actual offset of the forgery modulo 2. We Image Ground Truth Proposed ManTraNet [9], [45] Splicebuster [47] Noiseprint [10] Comprint [46] Noisesniffer [7] ZERO [8], [43] AdaCFA [5] Fig. 2: Visual results on the proposed dataset.…”

A New Benchmark and Model for Challenging Image Manipulation Detection

“…To accomplish this task, most of them utilize high-pass noise filters (Bayar and Stamm 2018;Li and Huang 2019) to suppress content information. Other approaches (Kwon et al 2022;Park et al 2018a;Mareen et al 2022) seek to identify compression inconsistencies in tampered images, as they assume that the compression QF's before and after manipulation differ. In addition to these two mainstream approaches, some researchers have directed their attention to camerabased artifacts, such as model fingerprints (Cozzolino and Verdoliva 2019;Cozzolino, Poggi, and Verdoliva 2015;Huh et al 2018;Mareen et al 2022).…”