Search

Differential privacy (DP) is a mathematical definition of privacy that can be widely applied when publishing data. DP has been recognized as a potential means of adhering to various privacy-related legal requirements. However, it can be difficult to reason about whether DP may be appropriate for a given context due to tensions that arise when it is brought from theory into practice. To aid policymaking around privacy concerns, we identify three categories of challenges to understanding DP along with associated questions that policymakers can ask about the potential deployment context to anticipate its impacts., Comment: This paper is accepted for publication in Policy Insights from the Behavioral and Brain Sciences (2024)

Differential privacy is a popular privacy-enhancing technology that has been deployed both in industry and government agencies. Unfortunately, existing explanations of differential privacy fail to set accurate privacy expectations for data subjects, which depend on the choice of deployment model. We design and evaluate new explanations of differential privacy for the local and central models, drawing inspiration from prior work explaining other privacy-enhancing technologies. We find that consequences-focused explanations in the style of privacy nutrition labels that lay out the implications of differential privacy are a promising approach for setting accurate privacy expectations. Further, we find that while process-focused explanations are not enough to set accurate privacy expectations, combining consequences-focused explanations with a brief description of how differential privacy works leads to greater trust.

Differential privacy (DP) has the potential to enable privacy-preserving analysis on sensitive data, but requires analysts to judiciously spend a limited ``privacy loss budget'' $\epsilon$ across queries. Analysts conducting exploratory analyses do not, however, know all queries in advance and seldom have DP expertise. Thus, they are limited in their ability to specify $\epsilon$ allotments across queries prior to an analysis. To support analysts in spending $\epsilon$ efficiently, we propose a new interactive analysis paradigm, Measure-Observe-Remeasure, where analysts ``measure'' the database with a limited amount of $\epsilon$, observe estimates and their errors, and remeasure with more $\epsilon$ as needed. We instantiate the paradigm in an interactive visualization interface which allows analysts to spend increasing amounts of $\epsilon$ under a total budget. To observe how analysts interact with the Measure-Observe-Remeasure paradigm via the interface, we conduct a user study that compares the utility of $\epsilon$ allocations and findings from sensitive data participants make to the allocations and findings expected of a rational agent who faces the same decision task. We find that participants are able to use the workflow relatively successfully, including using budget allocation strategies that maximize over half of the available utility stemming from $\epsilon$ allocation. Their loss in performance relative to a rational agent appears to be driven more by their inability to access information and report it than to allocate $\epsilon$., Comment: Published in IEEE Symposium on Security and Privacy (SP) 2024

Machine learning (ML) methods are proliferating in scientific research. However, the adoption of these methods has been accompanied by failures of validity, reproducibility, and generalizability. These failures can hinder scientific progress, lead to false consensus around invalid claims, and undermine the credibility of ML-based science. ML methods are often applied and fail in similar ways across disciplines. Motivated by this observation, our goal is to provide clear reporting standards for ML-based science. Drawing from an extensive review of past literature, we present the REFORMS checklist ($\textbf{Re}$porting Standards $\textbf{For}$ $\textbf{M}$achine Learning Based $\textbf{S}$cience). It consists of 32 questions and a paired set of guidelines. REFORMS was developed based on a consensus of 19 researchers across computer science, data science, mathematics, social sciences, and biomedical sciences. REFORMS can serve as a resource for researchers when designing and implementing a study, for referees when reviewing papers, and for journals when enforcing standards for transparency and reproducibility.

Differential privacy (DP) is a mathematical privacy notion increasingly deployed across government and industry. With DP, privacy protections are probabilistic: they are bounded by the privacy budget parameter, $\epsilon$. Prior work in health and computational science finds that people struggle to reason about probabilistic risks. Yet, communicating the implications of $\epsilon$ to people contributing their data is vital to avoiding privacy theater -- presenting meaningless privacy protection as meaningful -- and empowering more informed data-sharing decisions. Drawing on best practices in risk communication and usability, we develop three methods to convey probabilistic DP guarantees to end users: two that communicate odds and one offering concrete examples of DP outputs. We quantitatively evaluate these explanation methods in a vignette survey study ($n=963$) via three metrics: objective risk comprehension, subjective privacy understanding of DP guarantees, and self-efficacy. We find that odds-based explanation methods are more effective than (1) output-based methods and (2) state-of-the-art approaches that gloss over information about $\epsilon$. Further, when offered information about $\epsilon$, respondents are more willing to share their data than when presented with a state-of-the-art DP explanation; this willingness to share is sensitive to $\epsilon$ values: as privacy protections weaken, respondents are less likely to share data.

Recent arguments that machine learning (ML) is facing a reproducibility and replication crisis suggest that some published claims in ML research cannot be taken at face value. These concerns inspire analogies to the replication crisis affecting the social and medical sciences. They also inspire calls for the integration of statistical approaches to causal inference and predictive modeling. A deeper understanding of what reproducibility concerns in supervised ML research have in common with the replication crisis in experimental science puts the new concerns in perspective, and helps researchers avoid "the worst of both worlds," where ML researchers begin borrowing methodologies from explanatory modeling without understanding their limitations and vice versa. We contribute a comparative analysis of concerns about inductive learning that arise in causal attribution as exemplified in psychology versus predictive modeling as exemplified in ML. We identify themes that re-occur in reform discussions, like overreliance on asymptotic theory and non-credible beliefs about real-world data generating processes. We argue that in both fields, claims from learning are implied to generalize outside the specific environment studied (e.g., the input dataset or subject sample, modeling implementation, etc.) but are often impossible to refute due to undisclosed sources of variance in the learning pipeline. In particular, errors being acknowledged in ML expose cracks in long-held beliefs that optimizing predictive accuracy using huge datasets absolves one from having to consider a true data generating process or formally represent uncertainty in performance claims. We conclude by discussing risks that arise when sources of errors are misdiagnosed and the need to acknowledge the role of human inductive biases in learning and reform.

Organizations often collect private data and release aggregate statistics for the public's benefit. If no steps toward preserving privacy are taken, adversaries may use released statistics to deduce unauthorized information about the individuals described in the private dataset. Differentially private algorithms address this challenge by slightly perturbing underlying statistics with noise, thereby mathematically limiting the amount of information that may be deduced from each data release. Properly calibrating these algorithms -- and in turn the disclosure risk for people described in the dataset -- requires a data curator to choose a value for a privacy budget parameter, $\epsilon$. However, there is little formal guidance for choosing $\epsilon$, a task that requires reasoning about the probabilistic privacy-utility trade-off. Furthermore, choosing $\epsilon$ in the context of statistical inference requires reasoning about accuracy trade-offs in the presence of both measurement error and differential privacy (DP) noise. We present Visualizing Privacy (ViP), an interactive interface that visualizes relationships between $\epsilon$, accuracy, and disclosure risk to support setting and splitting $\epsilon$ among queries. As a user adjusts $\epsilon$, ViP dynamically updates visualizations depicting expected accuracy and risk. ViP also has an inference setting, allowing a user to reason about the impact of DP noise on statistical inferences. Finally, we present results of a study where 16 research practitioners with little to no DP background completed a set of tasks related to setting $\epsilon$ using both ViP and a control. We find that ViP helps participants more correctly answer questions related to judging the probability of where a DP-noised release is likely to fall and comparing between DP-noised and non-private confidence intervals., Comment: Forthcoming in Proceedings on Privacy Enhancing Technologies 2022

The computer science research community and the broader public have become increasingly aware of negative consequences of algorithmic systems. In response, the top-tier Neural Information Processing Systems (NeurIPS) conference for machine learning and artificial intelligence research required that authors include a statement of broader impact to reflect on potential positive and negative consequences of their work. We present the results of a qualitative thematic analysis of a sample of statements written for the 2020 conference. The themes we identify broadly fall into categories related to how consequences are expressed (e.g., valence, specificity, uncertainty), areas of impacts expressed (e.g., bias, the environment, labor, privacy), and researchers' recommendations for mitigating negative consequences in the future. In light of our results, we offer perspectives on how the broader impact statement can be implemented in future iterations to better align with potential goals., Comment: 12 pages, AAAI/ACM Conference on Artificial Intelligence, Ethics, and Society (AIES) 2021; added reference and link to dataset

Making conjectures about future consequences of a technology is an exercise in trying to reduce various forms of uncertainty. Both to produce and reason about these conjectures requires understanding their potential limitations. In other words, we need systematic ways of considering uncertainty associated with given conjectures for downstream consequences. In this work, we frame the task of considering future consequences as an anticipatory ethics problem, where the goal is to develop scenarios that reflect plausible outcomes and their ethical implications following a technology's introduction into society. In order to shed light on how various forms of uncertainty might inform how we reason about a resulting scenario, we provide a characterization of the types of uncertainty that arise in a potential scenario-building process., Comment: Part of the Navigating the Broader Impacts of AI Research Workshop at NeurIPS 2020