CSA-Graphs: A Privacy-Preserving Structural Dataset for Child Sexual Abuse Research

Carlos Caetano; Camila Laranjeira; Clara Ernesto; Artur Barros; João Macedo; Leo S. F. Ribeiro; Jefersson A. dos Santos; Sandra Avila

cs.CVApr 8, 2026

CSA-Graphs: A Privacy-Preserving Structural Dataset for Child Sexual Abuse Research

Authors: Carlos Caetano, Camila Laranjeira, Clara Ernesto, Artur Barros, João Macedo, Leo S. F. Ribeiro, Jefersson A. dos Santos, Sandra Avila

Organizations: Universidade Estadual de Campinas (UNICAMP), Brazil · Instituto Federal de Educação, Ciência e Tecnologia de Minas Gerais (IFMG), Brazil · Universidade de São Paulo (USP), Brazil · Universidade Federal de Minas Gerais (UFMG), Brazil ·

Abstract

Child Sexual Abuse Imagery (CSAI) classification is an important yet challenging problem for computer vision research due to the strict legal and ethical restrictions that prevent the public sharing of CSAI datasets. This limitation hinders reproducibility and slows progress in developing automated methods. In this work, we introduce CSA-Graphs, a privacy-preserving structural dataset. Instead of releasing the original images, we provide structural representations that remove explicit visual content while preserving contextual information. CSA-Graphs includes two complementary graph-based modalities: scene graphs describing object relationships and skeleton graphs encoding human pose. Experiments show that both representations retain useful information for classifying CSAI, and that combining them further improves performance. This dataset enables broader research on computer vision methods for child safety while respecting legal and ethical constraints.

Figures & tables

Panel of Figure 1 — Figure 1: Example of the graph-based representations used in this work. (a) Original image. (b) Skeleton pose graph obtained through human pose estimation. (c) Scene graph describing objects and their relationships in the scene. The graphs are overlaid on the original image for visualization purposes only.

Dataset	Year	Works	LEA	Description
PE1-A	2010	[ 12 ]	BR-FP	80GB seized hard disk including novel CSAI.
ULS-A	2011	[ 45 ]	EU	36,198 images (20,000 CSAI) classified as CSAI, porn, or world.
SCH-A	2014	[ 43 , 37 ]	EU	60,000 images (20,000 CSAI) classified as CSAI, adult, or world.
RealCSA	2017	[ 14 ]	ES-NC	5,000 images (2,500 CSAI) of juveniles.
VTR-A	2018	[ 48 ]	BR-FP	Seized hard disk comprising 58,974 images with 33,723 CSAI.
RCPD	2018	[ 30 , 23 , 9 , 47 , 4 , 10 , 24 ]	BR-FP	2,138 images (836 CSAI) with region-based labels for age, gender, and nudity levels.

Table 1: Datasets containing real CSAI, summarized by year of introduction, works citing them, LEA origin, size, and label scheme. Information compiled from Macedo et al. [31] . LEA acronyms: BR-FP (Brazilian Federal Police), EU (European Union), AU-FP (Australian Federal Police), ES-NC (Spanish National Cybersecurity Institute), MT-PF (Malta Police Force), US (United States).

	All	CSAI	Non-CSAI
Triplets	82,983	40,624	42,359
Object classes	147	–	–
Predicate classes	37	–	–
Avg. triplets / image	50.91	48.54	53.42
Avg. unique objects / image	26.73	26.29	27.20
Avg. unique predicates / image	10.08	10.18	9.98

Table 2: Statistics of the scene graphs in CSA-Graphs .

Panel of Figure 2 — Figure 2: Most frequent scene graph elements in CSA-Graphs.

	All	CSAI	Non-CSAI
Skeleton poses	2,135	1,065	1,070
Mean skeletons per image	1.31	1.29	1.33
Mean detected keypoints/skeleton	11.55	11.18	11.91
Mean keypoint confidence	0.669	0.650	0.688

Table 3: Statistics of the detected skeleton poses in CSA-Graphs.

Figure 3: Skeleton pose keypoint detection rate in CSA-Graphs.

Method	Acc. (%)	Precis. (%)	Recall (%)	F1 (%)
SG-baseline	$74.47$ $\pm~0.01$	76.96 $\pm~0.01$	$72.49$ $\pm~0.03$	$74.31$ $\pm~0.02$
Skl-baseline	$64.39$ $\pm~0.01$	$64.77$ $\pm~0.01$	$67.16$ $\pm~0.02$	$65.66$ $\pm~0.01$
Fusion-baseline	75.33 $\pm~0.01$	$76.38$ $\pm~0.01$	74.96 $\pm~0.01$	75.62 $\pm~0.01$

Table 4: Baseline performance on CSA-Graphs using different input modalities. SG-baseline operates on scene graphs, Skl-baseline uses skeleton, and Fusion-baseline combines both modalities. Results were obtained using 5-fold cross-validation.

Figure 4: Qualitative examples illustrating complementary nature of the representations. Top: a case where pose-based cues help the classification, while scene graphs fail. Bottom: a case where objects and predicates help the classification, while pose fails.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: Examples of skeleton pose estimation limitations in CSAI scenarios. The top row shows images with only a single individual, resulting in a single detected skeleton and limited relational pose information. The middle row presents close-up views of specific body regions, where reliable full-body pose estimation cannot be obtained. The bottom row illustrates cases where skeleton extraction is incomplete due to occlusions, truncation, or challenging viewpoints.

Figure 6: Skeleton pose keypoint statistics in CSA-Graphs.

Panel of Figure 7 — Figure 7: Confusion matrices of the baseline models evaluated in CSA-Graphs.

Prediction pattern	Samples	CSAI	Non-CSAI
Pose correct / Scene wrong	175	109	66
Scene correct / Pose wrong	350	137	213
Scene wrong / Fusion correct	143	86	57
Pose wrong / Fusion correct	372	162	210

Appendix

Table 5: Prediction disagreement and correction analysis between skeleton-based, scene graph, and fusion models.

Triplet	Count	Sum score	Mean score
(hair, on, head)	1270	804.478	0.633
(nose, on, face)	862	594.013	0.689
(woman, has, hair)	685	403.372	0.589
(woman, laying on, bed)	369	369.000	1.000
(girl, has, hair)	627	352.883	0.563
(eye, on, face)	505	314.005	0.622

Appendix

Table 6: Top triplets in CSAI samples ranked by summed confidence score.

Triplet	Count	Sum score	Mean score
(hair, on, head)	1569	1080.801	0.689
(nose, on, face)	1350	1019.411	0.755
(woman, has, hair)	998	563.683	0.565
(eye, on, face)	732	535.964	0.732
(ear, on, head)	693	500.922	0.723
(boy, wearing, shirt)	826	451.690	0.547

Appendix

Table 7: Top triplets in Non-CSAI samples ranked by summed confidence score.

Scene Graphs

See all 17 papers

Human-Centric Perception for Child Sexual Abuse Imagery
Mar 28, 2026Camila Laranjeira, João Macedo, Sandra Avila +2Keypoint Detection Human Pose Estimation
Universidade Federal de Minas Gerais (UFMG), Brazil. · the Federal Police of Brazil. · Universidade Estadual de Campinas (UNICAMP), Brazil. +1
Classifying by Proxy: Explainable and Reproducible Ensemble of Proxy Tasks for Child Sexual Abuse Imagery Classification
Jun 14, 2026Clara Ernesto, Carlos Caetano, Sandra Avila +3Explainable Artificial Intelligence Ensemble Learning
Instituto de Ciências Matemáticas e de Computação (ICMC), Universidade Estadual de São Paulo (USP), Brazil · Instituto de Computação (IC), Universidade Estadual de Campinas (UNICAMP), Brazil · Departamento de Ciência da Computação, Universidade Federal de Minas Gerais (UFMG), Brazil +1
Position: Preventing AI-Generated CSAM Necessitates New Approaches to AI Safety
Jun 9, 2026Neil Kale, Rebecca Portnoff, Pratiksha Thaker +5AI Safety Evaluation AI Safety
1Carnegie Mellon University · 2Thorn

Mar 28, 2026J/K move · Enter open · S save

Human-Centric Perception for Child Sexual Abuse Imagery

Camila Laranjeira, João Macedo, Sandra Avila, Fabrício Benevenuto+1

Universidade Federal de Minas Gerais (UFMG), Brazil. · the Federal Police of Brazil. · Universidade Estadual de Campinas (UNICAMP), Brazil.+1

Keypoint Detection Human Pose Estimation Object Detection

Open paper AskPDF arXiv

Law enforcement agencies and non-gonvernmental organizations handling reports of Child Sexual Abuse Imagery (CSAI) are overwhelmed by large volumes of data, requiring the aid of automation tools. However, defining sexual abuse in images of children is inherently challenging, encompassing sexually explicit activities and hints of sexuality conveyed by the individual's pose, or their attire. CSAI classification methods often rely on black-box approaches, targeting broad and abstract concepts such as pornography. Thus, our work is an in-depth exploration of tasks from the literature on Human-Centric Perception, across the domains of safe images, adult pornography, and CSAI, focusing on targets that enable more objective and explainable pipelines for CSAI classification in the future. We introduce the Body-Keypoint-Part Dataset (BKPD), gathering images of people from varying age groups and sexual explicitness to approximate the domain of CSAI, along with manually curated hierarchically structured labels for skeletal keypoints and bounding boxes for person and body parts, including head, chest, hip, and hands. We propose two methods, namely BKP-Association and YOLO-BKP, for simultaneous pose estimation and detection, with targets associated per individual for a comprehensive decomposed representation of each person. Our methods are benchmarked on COCO-Keypoints and COCO-HumanParts, as well as our human-centric dataset, achieving competitive results with models that jointly perform all tasks. Cross-domain ablation studies on BKPD and a case study on RCPD highlight the challenges posed by sexually explicit domains. Our study addresses previously unexplored targets in the CSAI domain, paving the way for novel research opportunities.

Abstract

Scene Graphs

Explore similar work

Human-Centric Perception for Child Sexual Abuse Imagery

Classifying by Proxy: Explainable and Reproducible Ensemble of Proxy Tasks for Child Sexual Abuse Imagery Classification

Position: Preventing AI-Generated CSAM Necessitates New Approaches to AI Safety