Shahnewaz Karim Sakib, Anindya Bijoy Das, Shibbir Ahmed. Proceedings of the 5th Workshop on Trustworthy NLP (TrustNLP 2025). 2025.

Information leakage is usually defined as the logarithmic increment in the adversary’s probability of correctly guessing the legitimate user’s private data or some arbitrary function of the private data when presented with the legitimate user’s publicly disclosed information. However, this definition of information leakage implicitly assumes that both the privacy mechanism and the prior probability of the original data are entirely known to the attacker. In reality, the assumption of complete knowledge of the privacy mechanism for an attacker is often impractical. The attacker can usually have access to only an approximate version of the correct privacy mechanism, computed from a limited set of the disclosed data, for which they can access the corresponding un-distorted data. In this scenario, the conventional definition of leakage no longer has an operational meaning. To address this problem, in this article, we propose novel meaningful information-theoretic metrics for information leakage when the attacker has incomplete information about the privacy mechanism—we call them average subjective leakage , average confidence boost , and average objective leakage , respectively. For the simplest, binary scenario, we demonstrate how to find an optimized privacy mechanism that minimizes the worst-case value of either of these leakages.

The ``Right to be Forgotten," as outlined in regulatory frameworks such as the General Data Protection Regulation (GDPR) and the California Consumer Privacy Act (CCPA), allows individuals to request the deletion of their personal data from deployed machine learning models. This provision ensures that individuals can maintain control over their personal information. In the digital health era, this right has become a critical concern for both patients and healthcare providers. To facilitate the effective removal of personal data from machine learning models, the concept of ``machine unlearning" has been introduced. This paper highlights the technical and ethical challenges associated with machine unlearning in digital healthcare. By examining current unlearning methodologies and their limitations, we propose a roadmap for future research and development in this field.

Traditional approaches to authenticated key establishment include the use of PKI or trusted third parties. While certificate deployment is sub-optimal for large-scale, low-cost applications, the use of trusted third parties is subject to human error and leaked credentials. For this context, co-location can be a valuable resource, and it is often exploited through common randomness harvesting techniques, but these, in turn, suffer from low achievable rates and usually from restrictive assumptions about the environment. Recent techniques for exploiting co-location are based on the notion of quality time and rely on sophisticated throttled clue-issuing mechanisms that allow a device with enough time to spend in the vicinity of the transmitter to find a secret key by collecting enough consecutive clues. By contrast, attackers are afforded only limited time to listen to, or interact with, the clue transmitter. Previous work in this direction deals solely with passive attackers and uses high-overhead information throttling mechanisms. This paper introduces the active attacker model for the quality-time paradigm and proposes a simple solution, a Zeroconf Key Establishment Protocol (ZeroProKeS). Additionally, the paper shows how to efficiently expand the proposed protocol to adhere to any customized information transfer function between legitimate users.

This paper analyzes the problem of estimating information leakage when the complete statistics of the privacy mechanism are not known, and the only available information consists of several input-output pairs obtained through interaction with the system or through some side channel. Several metrics, such as subjective leakage, objective leakage, and confidence boost, were introduced before for this purpose, but by design only work in a Bayesian framework. However, it is known that Bayesian inference can quickly become intractable if the domains of the involved variables are large. In this paper, we focus on this exact problem and propose a novel approach to perform an estimation of the leakage measures when the true knowledge of the privacy mechanism is beyond the reach of the user for a non-Bayesian framework using machine learning. Initially, we adapt the definition of leakage metrics to a non-Bayesian framework and derive their statistical bounds, and afterward, we evaluate the performance of those metrics via various experiments using Neural Networks, Random Forest Classifiers, and Support Vector Machines. We have also evaluated their performance on an image dataset to demonstrate the versatility of the metrics. Finally, we provide a comparative analysis between our proposed metrics and the metrics of the Bayesian framework.