Data collected from individuals is the fuel that drives the modern AI revolution. Intense recent discussions have focused on how to provide individuals control over when their data can and cannot be used---the EU's Right To Be Forgotten regulation is an example of this effort. Here we initiate a framework to study what to do when it is no longer permissible to deploy models derivative from certain individual user data. In particular, we formulate the problem of how to efficiently delete individual data from ML models that have been trained using this data. For many standard ML models, the only way to completely remove a person's data is to retrain the whole model from scratch on the remaining data. This is not practically feasible in many settings and we investigate ML models where data can be efficiently deleted

No abstract available.

As algorithms are increasingly used to make important decisions pertaining to individuals, algorithmic discrimination is becoming a prominent concern. The seminal work of Dwork et al. [ITCS 2012] introduced the notion of individual fairness (IF): given a task-specific similarity metric, every pair of similar individuals should receive similar outcomes. In this work, we study fairness when individuals have diverse preferences over the possible outcomes. We show that in such settings, individual fairness can be too restrictive: requiring individual fairness can lead to less-preferred outcomes for the very individuals that IF aims to protect (e.g. a protected minority group).

We introduce and study a new notion of preference-informed individual fairness (PIIF), a relaxation of individual fairness that allows for outcomes that deviate from IF, provided the deviations are in line with individuals' preferences. We show that PIIF can allow for solutions that are considerably more beneficial to individuals than the best IF solution. Motivated by fairness concerns in targeted advertising, we apply this new fairness notion to the multiple-task setting introduced by Dwork and Ilvento [ITCS 2019]. Finally, we demonstrate that in both settings, any convex objective over the outcomes subject to PIIF can be efficiently optimized for a rich class of individual preferences.

No abstract available.

Hardware as a root of trust presents interesting - and intriguing - alternative to secure multi-party computations. We survey several approaches to hardware-based security, such as hardware tokens, secure elements, and secure enclaves, with application towards privacy amplification via shuffling. We conclude with open research directions and problems.

In the 90s, software engineering shifted from packaged software and PCs to services and clouds, enabling distributed architectures that incorporate real-time feedback from users. In the process, digital systems became layers of technologies metricized under the authority of objective functions. These functions drive selection of software features, service integration, cloud usage, user interaction and growth, customer service, and environmental capture, among others. Whereas information systems focused on storage, processing and transport of information, and organizing knowledge "with associated risks of surveillance" contemporary systems leverage the knowledge they gather to not only understand the world, but also to optimize it, seeking maximum extraction of economic value through the capture and manipulation of people's activities and environments. The ability of these optimization systems to treat the world not as a static place to be known, but as one to sense and co-create, poses social risks and harms such as social sorting, mass manipulation, asymmetrical concentration of resources, majority dominance, and minority erasure. In the vocabulary of optimization, these harms arise due to choosing inadequate objective functions. During the talk, I will provide an account of what we mean by optimization systems, detail their externalities and make a proposition for Protective Optimization Technologies.

Multiple hypothesis testing, a situation when we wish to consider many hypotheses, is a core problem in statistical inference that arises in almost every scientific field. In this setting, controlling the false discovery rate (FDR), which is the expected proportion of type I error, is an important challenge for making meaningful inferences. In this talk, we consider a setting where an ordered (possibly infinite) sequence of hypotheses arrives in a stream, and for each hypothesis we observe a p-value along with a set of features specific to that hypothesis. The decision whether or not to reject the current hypothesis must be made immediately at each timestep, before the next hypothesis is observed.  We propose a new class of powerful online testing procedures, where the rejection thresholds are learned sequentially by incorporating contextual information and previous results. Any rule in this class controls online FDR under some standard assumptions. We will also discuss how our proposed procedures would lead to an increase of statistical power over a popular online testing procedure.

The General Data Protection Regulation (GDPR) and the California Consumer Privacy Act (CCPA) are the most consequential developments in information policy in a generation. In this talk, I will explain their strategic goals and the most interesting mechanisms these instruments use to create incentives and disincentives, to structurally strengthen some relationships while disadvantaging others, and to create privacy markets.

This talk will consider a paradigm where the public generally own and control their data. Here, mechanisms for purchasing data could create huge economic value (and share it fairly). But putting this into practice requires assigning value to data and efficiently matching the mechanism to providers. I will outline some research on this problem utilizing ideas from prediction markets, active learning, and differential privacy. We will see a few important directions for future work. I may also ramble briefly regarding software freedom, labor versus capital, etc.

Accountability is used often in describing computer-security mechanisms that complement preventive security, but it lacks a precise, agreed-upon definition.  Here, we argue for the need for accountability in computing in a variety of settings, categorize some of the many ways in which this term is used, and propose a punishment-focused view of "accountability."  We formalize our view in a utility-theoretic way and then use this to reason about accountability in computing systems.  We also survey mechanisms providing various senses of accountability as well as other approaches to reasoning about accountability-related properties. 

This is joint work with Joan Feigenbaum and Aaron Jaggard.