AnonShield: Scalable On-Premise Pseudonymization for CSIRT Vulnerability Data

Cristhian Kapelinski; Douglas Lautert; Beatriz Machado; Diego Kreutz; Isadora Garcia Ferrão

cs.CRApr 5, 2026

AnonShield: Scalable On-Premise Pseudonymization for CSIRT Vulnerability Data

Authors: Cristhian Kapelinski, Douglas Lautert, Beatriz Machado, Diego Kreutz, Isadora Garcia Ferrão

Organizations: Federal University of Pampa (UNIPAMPA) · Universit´e de Bretagne Occidentale (UBO)

Abstract

We present AnonShield, a high-throughput, on-premise pseudonymization system that combines GPU-accelerated NER, streaming processing, caching, and schema-aware configuration. Evaluated on datasets up to 550 MB (70,951 records), AnonShield reduces processing time from over 92 hours to under 10 minutes (up to 738x speedup) while achieving up to 94.2% F1-score and 96.7% recall. Our results show that scalable pseudonymization of vulnerability data is feasible without sacrificing analytical utility, enabling compliant data sharing in operational CSIRT environments.

Figures & tables

System	Main Technique	Target Domain	Type	File Formats	Gap in CSIRT context
Microsoft Presidio [ 15 ]	NLP + RegEx + extensible recognizers	Generic text	US	–	No CVE/CPE/hash recognizers; domain-irrelevant patterns produce FPs in vulnerability data
Google Cloud DLP [ 11 ]	NLP + FPE + PII classification	General data	US+S	.txt, .csv, .json, .pdf, .docx	Cloud endpoint conflicts with CSIRT data sovereignty policies; no cybersecurity-specific entities
Amazon Comprehend [ 4 ]	Cloud NLP for PII detection and redaction	Generic text	US	.txt	Cloud endpoint conflicts with CSIRT data sovereignty policies; no vulnerability scanner format support
IRI DarkShield [ 12 ]	Multi-source masking + REST API	Corporate data	US+S	.txt, .pdf, .xml, .json, .sql, images	Commercial licence limits adoption in academic and public-sector CSIRTs; no cybersecurity-specific entities
LogLicker [ 1 ]	RegEx + masking + remapping manifest	CloudTrail & Generic logs	US+S	.json, .txt	RegEx only; lacks contextual NER; high manual maintenance for custom logs
Anonip [ 9 ]	Bit masking of trailing IP bits via pipe	Web server logs	US	.log, .txt	Covers IP addresses only

Table I: Overview of anonymisation and pseudonymisation tools and their primary gap relative to the CSIRT vulnerability processing context. Data type: US = Unstructured; S = Structured. – in Formats: tool operates on in-memory strings with no native file I/O.

Figure 1: AnonShield pipeline. The same entity produces a consistent pseudonym across documents given the same secret key.

ID	Source	Format	Size	Access
D1	OpenVAS (VulnLab)	CSV/TXT/PDF/XML	9.29/11.61/33.32/34.11 MB	Public
D1C	Converted from D1	XLSX/DOCX/JSON/PDF(img)	3.34/7.58/60.72/1,479.56 MB	Public
D2	Tenable (CAIS/RNP)	CSV/JSON	419.72/550.54 MB	Private
D3	Synthetic (Mock)	CSV/JSON	247.45/444.56 MB	Public

Table II: Dataset summary.

Category	Examples
Network & Infrastructure	DNS (Bind), FTP (ProFTPD), OpenLDAP
Web Applications & APIs	Juice Shop, DVWA, GraphQL
Web Servers & Platforms	Apache, Nginx, Tomcat, WordPress
Databases	MySQL, PostgreSQL, MongoDB, Redis
Monitoring & Logging	Elasticsearch, Prometheus, Grafana
DevOps & CI/CD	Jenkins, GitLab, Nexus

Table III: VulnLab service categories.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Format	Version/Strategy	Mean (s)	Median (s)	Max (s)	CV	Speedup	Cohen’s $d$
XML	AnonLFI v2.0 (Baseline)	191.89	176.44	1717.71	0.96	$1.00\times$	–
	AnonShield_presidio	15.62	13.69	85.58	0.60	$12.28\times$	1.29 (Large)
	AnonShield_standalone	11.62	10.02	72.34	0.63	$\mathbf{16.51\times}$	1.32 (Large)
CSV	AnonLFI v2.0 (Baseline)	74.18	37.01	1442.08	1.75	$1.00\times$	–
	AnonLFI v1.0	37.92	20.41	577.96	1.55	$1.96\times$	0.35 (Small)
	AnonShield_presidio	10.65	9.08	60.62	0.53	$6.97\times$	0.67 (Medium)

Appendix

Table IV: Processing latency on D1 reports.

Figure 2: Scalability analysis on D1 PDF subset: (A) Linear scale, (B) Log-Log complexity (

\alpha\approx 1

for AnonShield), and (C) Throughput stability.

Format	Version/Strategy	Mean (s)	Max (s)	CV	Speedup	Cohen’s $d$
JSON	AnonLFI v2.0 (Baseline)	246.55	1861.01	1.16	$1.00\times$	–
	AnonShield_presidio	20.15	110.78	0.72	$12.23\times$	1.08 (Large)
	AnonShield_standalone	10.61	49.02	0.60	$\mathbf{23.24\times}$	1.12 (Large)
XLSX	AnonLFI v1.0 (Baseline)	35.51	320.73	1.37	$1.00\times$	–
	AnonLFI v2.0	60.30	596.95	1.52	$0.59\times$	$-0.34$ (Small)
	AnonShield_presidio	9.86	38.64	0.47	$3.60\times$	0.73 (Medium)

Appendix

Table V: Processing latency on D1-converted formats.

		CSV (419.72 MB)		JSON (550.54 MB)
Mode	Version / Strategy	Time (s)	KB/s	Time (s)	KB/s
No config	AnonLFI v2.0 (est.)	${\geq}437{,}335$	0.98	${\sim}334{,}472$	1.69
	AnonShield_presidio	$2{,}520.7\pm 68.0$	171	$985.7\pm 73.7$	575
	AnonShield_standalone	$\mathbf{588.5\pm 30.7}$	732	$\mathbf{453.1\pm 35.9}$	1,250
With config	AnonShield_presidio	$13.42\pm 0.13$	32,034	$18.88\pm 0.16$	29,855
With config	AnonShield_standalone	$\mathbf{12.55\pm 0.74}$	34,341	$\mathbf{18.03\pm 0.23}$	31,272

Appendix

Table VI: D2 (Operational) processing times (70,951 records).

		CSV (247.45 MB)		JSON (444.56 MB)
Mode	Version / Strategy	Time (s)	KB/s	Time (s)	KB/s
No config (GPU)	AnonLFI v2.0 (est.)	${\geq}257{,}835$	0.98	${\sim}270{,}086$	1.69
	AnonShield_presidio	$318.0\pm 5.1$	797	$339.6\pm 14.4$	1,343
	AnonShield_standalone	$\mathbf{73.0\pm 1.6}$	3,473	$\mathbf{172.1\pm 6.2}$	2,647
No config (CPU)	AnonShield_standalone	$\mathbf{481.5\pm 8.9}$	526	$\mathbf{881.9\pm 57.7}$	518
With config	AnonShield_presidio	$8.89\pm 0.06$	28,517	$21.11\pm 0.22$	21,571

Appendix

Table VII: D3 (Synthetic) processing times.

Version/Strategy	TP	FP	FN	Prec.	Rec.	F1
AnonLFI v1.0	106	224	459	32.1%	18.8%	23.7%
AnonLFI v2.0	279	68	407	80.4%	40.7%	54.0%
AnonShield_presidio	724	287	25	71.6%	96.7%	82.3%
AnonShield_filtered	724	64	25	91.9%	96.7%	94.2%
AnonShield_hybrid	724	64	25	91.9%	96.7%	94.2%
AnonShield_standalone	739	102	43	87.9%	94.5%	91.1%

Appendix

Table VIII: Accuracy on 67-record validation set (D1 OpenVAS CSV).

Cyber Threat Intelligence

Cyber threat intelligence turns evidence about adversaries, campaigns, and vulnerabilities into knowledge that helps defenders anticipate and respond to attacks.

See all 41 papers

VeriX-Anon: A Multi-Layered Framework for Mathematically Verifiable Outsourced Target-Driven Data Anonymization
Apr 14, 2026Miit Daga, Swarna Priya RamuPrivacy Auditing Verifiable Outsourced Computation
School of Computer Science Engineering and Information Systems, Vellore Institute of Technology, Vellore, 632014, Tamil Nadu, India
Subject-level Inference for Realistic Text Anonymization Evaluation
Apr 23, 2026Myeong Seok Oh, Dong-Yun Kim, Hanseok Oh +6Privacy Auditing
Tscientific, South Korea · Soongsil University, South Korea · Yonsei University, South Korea +1
Decomposing Memorization Reduction in Privacy-Preserving Fine-Tuning of SLMs for CSIRTs
Jun 26, 2026Cristhian Kapelinski, Diego KreutzPrivacy-Preserving Language Models Memorization in Language Models
Federal University of Pampa (UNIPAMPA), Alegrete, Brazil

Apr 14, 2026J/K move · Enter open · S save

VeriX-Anon: A Multi-Layered Framework for Mathematically Verifiable Outsourced Target-Driven Data Anonymization

Miit Daga, Swarna Priya Ramu

School of Computer Science Engineering and Information Systems, Vellore Institute of Technology, Vellore, 632014, Tamil Nadu, India

Privacy Auditing Verifiable Outsourced Computation Privacy-Preserving ML

Open paper AskPDF arXiv

Organisations increasingly outsource privacy-sensitive data transformations to cloud providers, yet no practical mechanism lets the data owner verify that the contracted algorithm was faithfully executed. VeriX-Anon is a multi-layered verification framework for outsourced Target-Driven k-anonymization combining three orthogonal mechanisms: deterministic verification via Merkle-style hashing of an Authenticated Decision Tree, probabilistic verification via Boundary Sentinels and exact-duplicate Twins with cryptographic identifiers, and utility-based verification via Explainable AI fingerprinting that compares SHAP value distributions before and after anonymization using the Wasserstein distance. Across seven cross-domain datasets and four cloud profiles (28 scenarios), against Lazy (drops records), Dumb (fake hash), and Approximate (valid hash) adversaries, VeriX-Anon detects 25 of 28 deviations under a fixed threshold and 27 of 28 once the threshold is calibrated per dataset, with no false alarms. No single layer achieved this alone. The XAI layer was the only mechanism that caught the Approximate adversary, succeeding on six of seven datasets and missing only a high-dimensional case where honest generalization shifts SHAP as much as the attack. Target-Driven anonymization preserved significantly more utility than blind splitting, with mean F1 gaps of +0.058 to +0.362 and Wilcoxon p <= 0.001 on six of seven datasets. Client-side verification completes under one second at one million rows. The threat model covers three empirically evaluated profiles and one theoretical Informed Attacker unable to defeat the cryptographic salt. Sentinel evasion probability ranges from near-zero to 0.82 for the most imbalanced data, which the twin layer offsets in every scenario.

Abstract

Cyber Threat Intelligence

Explore similar work

VeriX-Anon: A Multi-Layered Framework for Mathematically Verifiable Outsourced Target-Driven Data Anonymization

Subject-level Inference for Realistic Text Anonymization Evaluation

Decomposing Memorization Reduction in Privacy-Preserving Fine-Tuning of SLMs for CSIRTs