Sensitive data detection

This repository contains the dataset used for sensitive data detection as described in the paper "Exploiting Large Language Models to Train Automatic Detectors of Sensitive Data" (currently submitted for review in the IRCDL 2024 Conference...). As described in the paper, the dataset has been generated with the OpenLLaMa model [1], and as such is licensed permissively under the Apache 2.0 license: we decided to distribute it freely to encourage further research in this field.

Format

The dataset is shaped as a list of entries: each top-level entry is a couple (text, label), where:

• text holds the content of the generated document;
• label is a list of entries, in with the key is a label specifying the category of sensitive data, the value is a list of pairs, each pair indicating the beginning and ending character index of the sensitive span.

Categories and labels

The labels used in the dataset follows with a brief description:

• [DATI_SALUTE]: This category encompasses information pertaining to the physical and mental well-being of individuals. It includes details regarding existing diagnoses, medical conditions, and disabilities.
• [DATI_POLITICA]: The political category involves information related to an individual's political beliefs. This encompasses their political orientation, specific party affiliation, as well as membership in work unions or similar organizations.
• [DATI_SESSUALITA]: Within the realm of sexuality, we consider information concerning an individual's sexual orientation, habits, and gender identity.
• [DATI_GIUDIZIARI]: The judicial category encompasses data associated with legal matters, such as offenses, crimes, charges, pending criminal proceedings, accusations, and trial proceedings involving an individual.
• [DATI_RELIGIONE]: Information regarding an individual's philosophical and religious beliefs and affiliations falls within the scope of the philosophy category.
• [DATI_ETNIA]: The ethnic category pertains to an individual's ethnic origin and heritage.

Document types and count

Data category	Doc title	Doc count	Avg char count
Mix (Sensitive)	Email	233	1072.91
	Newspaper article	148	1004.84
	Curriculum Vitae	109	1241.66
	TOT	490
Health & Sexuality	Psychiatric report	83	1638.61
	Medical prescription	71	1372.89
	Medical records	62	1476.06
	Psychological evaluation	53	1899.11
	Certification of invalidity	20	1419.95
	Biopsy results	21	1123.29
	Eye test report	15	1455.33
	Surgery report	15	1475.53
	Blood tests	10	1437.5
	Certificate of civil union	20	1116.95
	TOT	370
Judicial	Denunciation report	26	1225.58
	Police identikit	25	1097.16
	Criminal record	24	1305.08
	Arrest report	19	1234.16
	Notice of investigation	24	1806.88
	Criminal judgement	22	1403.82
	Notice of conclusion of preliminary investigations	19	1610.89
	Certificate of pending charges	14	1268.14
	Precautionary measures	18	1281.28
	TOT	191
Politic	Political endorsement	38	735.39
	Union card	30	1066.77
	Party card	28	1067.14
	TOT	96
Philosophical	Philosophical endorsement	68	1028.49
	Baptismal certificate	32	783.47
	Certificate of participation to religious group	32	682.59
	TOT	132
Ethnic	DNA analysis report	37	1804.86
	Ancestry analysis report	31	1824.16
	Birth certificate	36	435.61
	Genealogical tree report	30	1444.67
	TOT	134
Other (Non sensitive)	Scientific paper	106	3038.55
	Advertising flyer for event	92	992.42
	Scientific publications report	88	2576.69
	Marriage certificate	66	829.39
	Advertising flyer	18	1035.83
	Company invoice	17	1203.71
	Services and products catalogue	15	2122.2
	Financial report by corporate	14	2215.5
	Commercial report	13	2128.92
	City travel guide	12	1598.75
	Tax declaration	12	1476.58
	Cooking recipe	11	1168.64
	Corporate memo	9	1277.67
	Company balance sheet	8	1443.12
	Book review	7	767
	Wikipedia extract	150	2430.3
	TOT	638

References

[1] Geng, Xinyang and Liu, Hao, OpenLLaMA: An Open Reproduction of LLaMA, https://github.com/openlm-research/open_llama