Local Data Debiasing for Fairness Based on Generative Adversarial Training

Abstract

The widespread use of automated decision processes in many areas of our society raises serious ethical issues with respect to the fairness of the process and the possible resulting discrimination. To solve this issue, we propose a novel adversarial training approach called GANSan for learning a sanitizer whose objective is to prevent the possibility of any discrimination (i.e., direct and indirect) based on a sensitive attribute by removing the attribute itself as well as the existing correlations with the remaining attributes. Our method GANSan is partially inspired by the powerful framework of generative adversarial networks (in particular Cycle-GANs), which offers a flexible way to learn a distribution empirically or to translate between two different distributions. In contrast to prior work, one of the strengths of our approach is that the sanitization is performed in the same space as the original data by only modifying the other attributes as little as possible, thus preserving the interpretability of the sanitized data. Consequently, once the sanitizer is trained, it can be applied to new data locally by an individual on their profile before releasing it. Finally, experiments on real datasets demonstrate the effectiveness of the approach as well as the achievable trade-off between fairness and utility.

Record DNA: reconceptualising digital records as the future evidence base

“…The need for content with data quality, reliability and truthful trusted production, as opposed to the production of 'flawed' records, was highlighted with the dangers of 'fairwashing' highlighted. This is a conversation that has been growing in the technology sphere in terms of the risks of rationalising the use and explanations of the applications of AI/algorithms (Aïvodji et al 2019). For some, the provenance, including data provenance, and chain of custody or data lineage were the imperative elements to deliver recordness.…”

Learning to Distill Graph Neural Networks

“…The basic idea of such methods is to use gradient-based optimization to generate some samples of optimal representation in each target class [18], [19], [22], [37]. 2) In a black-box setting, since the parameters of the victim model cannot be accessed [2], it is impossible to generate images of the target class through gradient-based optimization directly. Therefore, some works choose to use auxiliary data similar to the original data to steal the knowledge of the API [12], [43].…”

Deep Learning Meets Deep Democracy: Deliberative Governance and Responsible Innovation in Artificial Intelligence

“…Common themes identified in the literature on expert opacity include so-called epistemic vices, such as AI “gullibility,” “dogmatism,” and “automation bias” (Tsamados et al, 2021). Expert opacity can also arise inadvertently through attempts at disclosure and transparency that overwhelm citizens on account of the sheer volume and complexity of information made available to them (Ananny & Crawford, 2018), though here it should be noted that any intentional obfuscation by such disclosure “overload” would rather constitute an element of strategic opacity (Aïvodji, Arai, Fortineau, Gambs, Hara, & Tapp, 2019).…”

“…This is because such efforts do not take place in a social vacuum but in specific cultural and organizational settings (Felzmann, Villaronga, Lutz, & Tamò-Larrieux, 2019; Kemper & Kolkman, 2019; Miller, 2019), meaning they are performative and may have unintended consequences and downsides (Albu & Flyverbom, 2019). This is the case, for example, in attempts at AI explicability that actually serve to obfuscate further through disclosure (Aïvodji et al, 2019; Ananny & Crawford, 2018).…”