Impact of Dataset Properties on Membership Inference Vulnerability of Deep Transfer Learning
Fuente:
arXiv
Salvato in:
| Autori principali: | , , , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2024
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866910007856988160 |
|---|---|
| author | Tobaben, Marlon Ito, Hibiki Jälkö, Joonas He, Yuan Honkela, Antti |
| author_facet | Tobaben, Marlon Ito, Hibiki Jälkö, Joonas He, Yuan Honkela, Antti |
| contents | Membership inference attacks (MIAs) are used to test practical privacy of machine learning models. MIAs complement formal guarantees from differential privacy (DP) under a more realistic adversary model. We analyse MIA vulnerability of fine-tuned neural networks both empirically and theoretically, the latter using a simplified model of fine-tuning. We show that the vulnerability of non-DP models when measured as the attacker advantage at a fixed false positive rate reduces according to a simple power law as the number of examples per class increases. A similar power-law applies even for the most vulnerable points, but the dataset size needed for adequate protection of the most vulnerable points is very large. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2402_06674 |
| institution | arXiv |
| publishDate | 2024 |
| record_format | arxiv |
| spellingShingle | Impact of Dataset Properties on Membership Inference Vulnerability of Deep Transfer Learning Tobaben, Marlon Ito, Hibiki Jälkö, Joonas He, Yuan Honkela, Antti Cryptography and Security Machine Learning Membership inference attacks (MIAs) are used to test practical privacy of machine learning models. MIAs complement formal guarantees from differential privacy (DP) under a more realistic adversary model. We analyse MIA vulnerability of fine-tuned neural networks both empirically and theoretically, the latter using a simplified model of fine-tuning. We show that the vulnerability of non-DP models when measured as the attacker advantage at a fixed false positive rate reduces according to a simple power law as the number of examples per class increases. A similar power-law applies even for the most vulnerable points, but the dataset size needed for adequate protection of the most vulnerable points is very large. |
| title | Impact of Dataset Properties on Membership Inference Vulnerability of Deep Transfer Learning |
| topic | Cryptography and Security Machine Learning |
| url | https://arxiv.org/abs/2402.06674 |