Events

In this talk I shall outline a summary of our five year long initiative studying the temporal dynamics of various human language-like entities over the social media. Some of the topics that I plan to cover are (a)  how opinion conflicts could be effectively used for incivility detection in Twitter [CSCW 2018], (b) how word borrowings can be automatically identified from social signals [EMNLP 2017] and (c)  how hashtags in Twitter form compounds like natural language words (e.g., #Wikipedia+#Blackout=#WikipediaBlackout) that become way more popular than the individual constituent hashtags [CSCW 2016, Honorable Mention].

-

In the early 1990s cryptography went into a foundational crisis when efficient quantum algorithms were discovered which could break almost all public key encryption schemes known at the time. Since then, a enormous research effort has been invested into basing public key cryptography, and secure computation in general, on problems which are conjectured to be hard even for quantum computers. While this research program has been resoundingly successful, even leading up the way to cryptographic milestones such as fully homomorphic encryption, there are still important cryptographic primitives for which no post-quantum secure protocols are known. Until very recently, one such primitive was 2-message oblivious transfer, a fundamental primitive in the field of secure two- and multi-party computation. I will discuss a novel construction of this primitive from the Learning With Errors (LWE) assumption, a lattice-based problem which is known to be as hard as worst-case lattice problems and conjectured to be post-quantum secure. The security of our construction relies on a fundamental Fourier-analytic property of lattices, namely the transference principle: Either a lattice or its dual must have short vectors.

Sophisticated mobile computing, sensing and recording devices like smartphones, smartwatches, and wearable cameras are carried by their users virtually around the clock, blurring the distinction between the online and offline worlds. While these devices enable transformative new applications and services, they also introduce entirely new threats to users' privacy, because they can capture a complete record of the user's location, online and offline activities, and social encounters, including an audiovisual record. Such a record of users' personal information is highly sensitive and is subject to numerous privacy risks. In this thesis, we have investigated and built systems to mitigate two such privacy risks: 1) privacy risks due to ubiquitous digital capture, where bystanders may inadvertently be captured in photos and videos recorded by other nearby users, 2) privacy risks to users' personal information introduced by a popular class of apps called `mobile social apps'. In this thesis, we present two systems, called I-Pic and EnCore, built to mitigate these two privacy risks.

Both systems aim to put the users back in control of what personal information is being collected and shared, while still enabling innovative new applications. We built working prototypes of both systems and evaluated them through actual user deployments. Overall we demonstrate that it is possible to achieve privacy-compliant digital capture and it is possible to build privacy-compliant mobile social apps, while preserving their intended functionality and ease-of-use. Furthermore, we also explore how the two solutions can be merged into a powerful combination, one which could enable novel workflows for specifying privacy preferences in image capture that do not currently exist.

Many classification tasks have no objective ground truth. Examples include: which content or explanation is "better" according to some community? is this comment toxic? what is the political leaning of this news article? The traditional modeling approach assumes each item has an objective true state that is perceived by humans with some random error. It fails to account for the fact that people have greater agreement on some items than others. I will describe an alternative model where the true state is a distribution over labels that raters from a specified population would assign to an item. This leads to information gain (mutual information) as a theoretically justified and computationally tractable measure of a classifier's quality, and an intuitive interpretation of information gain in terms of the sample size for a survey that would yield the same expected error rate.