SurFITR: A Dataset for Surveillance Image Forgery Detection and Localisation

Qizhou Wang; Guansong Pang; Christopher Leckie

cs.CVApr 8, 2026

SurFITR: A Dataset for Surveillance Image Forgery Detection and Localisation

Authors: Qizhou Wang, Guansong Pang, Christopher Leckie

Organizations: The University of Melbourne, Parkville, Australia · Singapore Management University, Singapore

Abstract

We present the Surveillance Forgery Image Test Range (SurFITR), a dataset for surveillance-style image forgery detection and localisation, in response to recent advances in open-access image generation models that raise concerns about falsifying visual evidence. Existing forgery models, trained on datasets with full-image synthesis or large manipulated regions in object-centric images, struggle to generalise to surveillance scenarios. This is because tampering in surveillance imagery is typically localised and subtle, occurring in scenes with varied viewpoints, small or occluded subjects, and lower visual quality. To address this gap, SurFITR provides a large collection of forensically valuable imagery generated via a multimodal LLM-powered pipeline, enabling semantically aware, fine-grained editing across diverse surveillance scenes. It contains over 137k tampered images with varying resolutions and edit types, generated using multiple image editing models. Extensive experiments show that existing detectors degrade significantly on SurFITR, while training on SurFITR yields substantial improvements in both in-domain and cross-domain performance. SurFITR is publicly available on GitHub.

Figures & tables

Figure 1. Visualisations from SurFITR showing realistic, fine grained tampering across diverse surveillance scenes. Top row: original images, middle row: tampered images (yellow boxes indicate manipulated regions), bottom row: edited pixel masks.

Figure 2. Comparison between SurFITR and prior datasets illustrating their differences. Red indicates manipulated regions, while absence denotes full generation or missing masks. Here, the SurFITR image removes the cashier, though the register was attended in the original image.

Figure 3. Overview of SurFITR. (a) Dataset structure illustrating the combinational design across source corpora, tampering types, entity types, and generation models. (b) Key statistics, including data distribution, edit types, manipulation scale, and generation models.

Figure 4. Overview of the SurFITR MLLM-powered tampering generation and verification pipeline. MLLMs are used to select forensically valuable frames from source corpora and, with assistance from grounding models, generate semantically and visually grounded tampering instructions. Localised, mask-guided manipulation is then applied, followed by verification to ensure tampering quality.

Task	Method	UCF		NTU		Shanghai		CUHK		UCSD 1		UCSD 2		Avg.		$\Delta$ vs. Verified
Task	Method	AUROC	F1	AUROC	F1	AUROC	F1	AUROC	F1	AUROC	F1	AUROC	F1	AUROC	F1	AUROC	F1
Det.	PSCC-Net	0.315	0.300	0.424	0.463	0.577	0.536	0.666	0.668	0.937	0.007	0.800	0.030	0.620	0.334	$+$ 0.011	$+$ 0.003
	MVSSNet	0.495	0.667	0.470	0.667	0.448	0.667	0.436	0.667	0.546	0.667	0.526	0.667	0.487	0.667	$-$ 0.035	0.000
	TruFor	0.491	0.575	0.559	0.600	0.430	0.666	0.420	0.667	0.461	0.667	0.625	0.396	0.498	0.595	$-$ 0.018	$-$ 0.003
	HiFi-Net	0.537	0.000	0.504	0.001	0.546	0.039	0.277	0.000	0.661	0.000	0.944	0.000	0.578	0.007	$-$ 0.002	$-$ 0.001
	DRCT-2M	0.701	0.049	0.524	0.405	0.790	0.301	0.527	0.296	0.796	0.000	0.630	0.000	0.662	0.175	$-$ 0.030	$+$ 0.010

Table 3. Zero-shot detection and localisation on SurFITR-Base;

\Delta

denotes the difference from the manually verified set.

Task	Train Set	Method		UCF		NTU		Shanghai		CUHK		UCSD 1		UCSD 2		Avg
Task	Train Set	Method		AUROC	F 1	AUROC	F 1	AUROC	F 1	AUROC	F 1	AUROC	F 1	AUROC	F 1	AUROC	F 1
Det.	Base (all)	PSCC-Net	ft	0.991	0.927	0.933	0.846	1.000	0.988	0.999	0.814	1.000	0.298	1.000	0.731	0.987	0.767
			$\Delta$	0.676	0.627	0.509	0.382	0.422	0.452	0.333	0.146	0.062	0.292	0.199	0.701	0.367	0.433
		TruFor	ft	0.823	0.577	0.432	0.417	0.786	0.199	0.852	0.809	0.990	0.560	0.928	0.562	0.802	0.521
			$\Delta$	0.326	0.041	-0.130	-0.153	0.344	-0.463	0.431	0.142	0.545	-0.108	0.307	0.559	0.304	0.003
		HiFi-Net	ft	0.998	0.986	0.961	0.877	1.000	0.985	0.996	0.991	1.000	1.000	1.000	1.000	0.993	0.973

Table 5. Finetuned results on SurFITR Base.

\Delta

: gain over zero-shot; green: improvement; red: degradation.

		Detection				Localisation
		Base (all)		Base (UCF)		Base (all)		Base (UCF)
Method		AUROC	F 1	AUROC	F 1	P-IoU	P-F 1	P-IoU	P-F 1
PSCC-Net	ft	0.952	0.575	0.793	0.548	0.058	0.083	0.018	0.028
	$\Delta$	0.286	0.191	0.128	0.164	0.031	0.039	-0.010	-0.015
TruFor	ft	0.785	0.502	0.655	0.351	0.115	0.139	0.081	0.100
	$\Delta$	0.257	-0.118	0.127	-0.269	0.047	0.052	0.013	0.014

Table 6. Avg. Finetuned performance on SurFITR-Transfer.

Image Forgery Detection

Image Forgery Detection studies how to identify manipulated or synthetic images and locate the visual evidence of tampering.

See all 126 papers

SARIF: Segment Anything for Robust Image Forensics
Jun 19, 2026Dong-Hyun Moon, Ju-Hyeon Nam, Sang-Chul LeeDigital Image Forensics Domain Generalization
Department of Electrical and Computer Engineering, Inha University, 100 Inha-ro, Michuhol-gu, Incheon, 22212, Republic of Korea
Learning Universal Features for Generalizable Image Forgery Localization
Apr 10, 2025Hengrun Zhao, Yunzhi Zhuge, Yifan Wang +3Image Forgery Detection Image Forgery Localization
Dalian University of Technology · Johns Hopkins University
Order within Chaos: Capturing Intrinsic Energy Anomalies for AI-Manipulated Image Forgery Localization
Jun 1, 2026Yiming Wang, Baiqi Wu, Qingming Li +3Image Forgery Detection Digital Image Forensics
Zhejiang University, Hangzhou, China.

Jun 19, 2026J/K move · Enter open · S save

SARIF: Segment Anything for Robust Image Forensics

Dong-Hyun Moon, Ju-Hyeon Nam, Sang-Chul Lee

Department of Electrical and Computer Engineering, Inha University, 100 Inha-ro, Michuhol-gu, Incheon, 22212, Republic of Korea

Digital Image Forensics Domain Generalization Image Forgery Localization

Open paper AskPDF arXiv

Image forgery localization remains challenging due to diverse manipulation techniques and distribution shifts. Existing forgery localization models achieve high accuracy on benchmarks but often struggle with cross-domain generalization and robustness. In this paper, we propose SARIF (Segment Anything for Robust Image Forensics), a framework that leverages the Segment Anything Model (SAM), which has a promptable architecture and strong generalization ability. SARIF introduces a feedback-guided mask decoder and a dual-encoder design that extracts forgery-specific information to capture forensic traces while exploiting the SAM architecture. To localize manipulated regions, we design a block-wise prompting mechanism that derives forgery-specific cues from residual features between an adapted encoder and its frozen counterpart. These features are fused with the previous mask prompt to drive a feedback-based mask refinement process, enabling automatic forgery segmentation without manual input. Extensive experiments on standard forgery-localization benchmarks show that SARIF achieves strong average cross-dataset performance and robustness to common image corruptions.

Abstract

Image Forgery Detection

Explore similar work

SARIF: Segment Anything for Robust Image Forensics

Learning Universal Features for Generalizable Image Forgery Localization

Order within Chaos: Capturing Intrinsic Energy Anomalies for AI-Manipulated Image Forgery Localization