Alexander Pugantsov
@pugantsov.com
Researcher @ CNR-ISTI; Machine learning under distribution shift; Previously @unipd.bsky.social, @irglasgow.bsky.social.
pugantsov.com
pugantsov.com
3. Attribute unawareness does not mean privacy.
Aggregate demographic estimates can leak enough information to infer individual sensitive attributes with high accuracy (macro-F1 > 0.9 in many settings), even when the attacker has relatively limited resources.
Aggregate demographic estimates can leak enough information to infer individual sensitive attributes with high accuracy (macro-F1 > 0.9 in many settings), even when the attacker has relatively limited resources.
September 27, 2026 at 3:01 PM
3. Attribute unawareness does not mean privacy.
Aggregate demographic estimates can leak enough information to infer individual sensitive attributes with high accuracy (macro-F1 > 0.9 in many settings), even when the attacker has relatively limited resources.
Aggregate demographic estimates can leak enough information to infer individual sensitive attributes with high accuracy (macro-F1 > 0.9 in many settings), even when the attacker has relatively limited resources.
2. Better demographic estimation does not necessarily mean better fairness estimation.
Downstream fairness assessment introduces additional complexity, and methods that perform well for direct demographic estimation can become much less reliable once used in multi-stage fairness evaluations.
Downstream fairness assessment introduces additional complexity, and methods that perform well for direct demographic estimation can become much less reliable once used in multi-stage fairness evaluations.
September 27, 2026 at 3:01 PM
2. Better demographic estimation does not necessarily mean better fairness estimation.
Downstream fairness assessment introduces additional complexity, and methods that perform well for direct demographic estimation can become much less reliable once used in multi-stage fairness evaluations.
Downstream fairness assessment introduces additional complexity, and methods that perform well for direct demographic estimation can become much less reliable once used in multi-stage fairness evaluations.
1. Demographic composition can be estimated surprisingly well without individual-level sensitive attributes: advanced methods reduced diversity estimation error by up to 81%.
However, estimation becomes much harder as demographic granularity and the complexity of dataset shift increase.
However, estimation becomes much harder as demographic granularity and the complexity of dataset shift increase.
September 27, 2026 at 3:01 PM
1. Demographic composition can be estimated surprisingly well without individual-level sensitive attributes: advanced methods reduced diversity estimation error by up to 81%.
However, estimation becomes much harder as demographic granularity and the complexity of dataset shift increase.
However, estimation becomes much harder as demographic granularity and the complexity of dataset shift increase.
Excited to share that our paper PAIRE is now online in Information Processing & Management!
We introduce a benchmark for estimating dataset diversity and model fairness without individual-level sensitive attributes, and examine the accuracy, robustness, and privacy risks of these estimates.
We introduce a benchmark for estimating dataset diversity and model fairness without individual-level sensitive attributes, and examine the accuracy, robustness, and privacy risks of these estimates.
September 27, 2026 at 3:01 PM
Excited to share that our paper PAIRE is now online in Information Processing & Management!
We introduce a benchmark for estimating dataset diversity and model fairness without individual-level sensitive attributes, and examine the accuracy, robustness, and privacy risks of these estimates.
We introduce a benchmark for estimating dataset diversity and model fairness without individual-level sensitive attributes, and examine the accuracy, robustness, and privacy risks of these estimates.