PMC

A natural language data de-identification and sharing pipeline from the publisher (left) to the attacker (right).

Performance for the publisher’s de-identification model as a function of the number of documents provided for training.

Decision making process of the publisher and corresponding strategies of the attacker in the Mid-high case when the attacker pays the penalty back to the publisher.

Decision making process of the publisher and corresponding strategies of the attacker in the Mid-high case when the attacker’s pays the penalty forward to a third party.

The influence of the publisher’s training dataset size on the attacker’s (a) precision, (b) recall, and (c) f-measure.

Experimental design for performance and cost model evaluations of the publisher and the attacker. Note that a specific experiment consists of the publisher and attacker choosing one level of training each.

A depiction of the traditional view on natural language de-identification (left) and an augmented view that accounts for potential attackers (right) and translation of traditional information retrieval measures into economic factors.

Sensitivity of attacker’s and publisher’s payoffs to the attacker’s value per true positive (v) and loss per false positive (l_a) for four policies when attacker’s annotation cost per EMR (c_a) is $4. The Mid-high case (v = $0.5, l_a = $0.5) is circled.

Sensitivity of attacker’s and publisher’s payoffs to the attacker’s value per true positive (v) and loss per false positive (l_a) for policies Traditional, Safe-forward, Attack-forward and Attack-back when the attacker’s annotation cost per EMR (c_a) is $1. The result for the Low case (v = $0.1, l_a = $0.3) is circled in each figure.