Giulia C. Fanti

Building a RAPPOR with the Unknown: Privacy-Preserving Learning of Associations and Data Dictionaries

Abstract Techniques based on randomized response enable the collection of potentially sensitive data from clients in a privacy-preserving manner with strong local differential privacy guarantees. A recent such technology, RAPPOR [12], enables estimation of the marginal frequencies of a set of strings via privacy-preserving crowdsourcing. However, this original estimation process relies on a known dictionary of possible strings; in practice, this dictionary can be extremely large and/or unknown. In this paper, we propose a novel decoding algorithm for the RAPPOR mechanism that enables the estimation of “unknown unknowns,” i.e., strings we do not know we should be estimating. To enable learning without explicit dictionary knowledge, we develop methodology for estimating the joint distribution of multiple variables collected with RAPPOR. Our contributions are not RAPPOR-specific, and can be generalized to other local differential privacy mechanisms for learning distributions of string-valued random variables.

Spy vs. Spy: Rumor Source Obfuscation

Anonymous messaging platforms, such as Secret, Yik Yak and Whisper, have emerged as important social media for sharing ones thoughts without the fear of being judged by friends, family, or the public. Further, such anonymous platforms are crucial in nations with authoritarian governments; the right to free expression and sometimes the personal safety of the author of the message depend on anonymity. Whether for fear of judgment or personal endangerment, it is crucial to keep anonymous the identity of the user who initially posted a sensitive message. In this paper, we consider an adversary who observes a snapshot of the spread of a message at a certain time. Recent advances in rumor source detection shows that the existing messaging protocols are vulnerable against such an adversary. We introduce a novel messaging protocol, which we call adaptive diffusion, and show that it spreads the messages fast and achieves a perfect obfuscation of the source when the underlying contact network is an infinite regular tree: all users with the message are nearly equally likely to have been the origin of the message. Experiments on a sampled Facebook network show that it effectively hides the location of the source even when the graph is finite, irregular and has cycles.

Circulant structures and graph signal processing

Linear shift-invariant processing of graph signals rests on circulant graphs and filters. The spatial features of circulant structures also permit shift-varying operations such as sampling. Their spectral features-as described by their Graph Fourier Transform profiles-enable novel multiscale signal processing systems and methods. To extend the reach of circulant structures, we present a method to decompose an arbitrary graph or filter into a combination of circulant structures. Our decomposition is analogous to resolving a linear time-varying system into a bank of linear time-invariant systems. As an application, we perform multiscale decomposition on temperature data spanning the continental United States.

Critically-sampled perfect-reconstruction spline-wavelet filterbanks for graph signals

Inspired by first-order spline wavelets in classical signal processing, we introduce two-channel (low-pass and high-pass), critically-sampled, perfect-reconstruction filterbanks for signals defined on circulant graphs, which accommodate linear shift-invariant filtering. We then generalize to filters that process signals defined on noncirculant graphs. We apply these filters, which can be tuned to approximate desired frequency responses, to signals defined on synthetic graphs and examine their performance.

One-Way Private Media Search on Public Databases: The Role of Signal Processing

Automated media classification is becoming increasingly common in areas ranging from mobile location recognition to surveillance systems to automated annotation. While these tools can add great value to the public sphere, media searches often process private information; in such situations, it is important to protect the interests of one or both parties. Much attention has been given to the scenario where both the server and the client wish to keep their data secret, but comparatively little work has been done on searches in which only the clients data is sensitive. Nonetheless, there is great potential for applications involving private searches on public databases like Google Images, Flickr, or Wanted Persons directories put forth by various police agencies. In this article, we make the case that one-way private media search is an important and practically viable direction for future research. We will introduce readers to some basic one-way privacy tools and present a case study outlining the design of a private audio search tool on a public database. This case study serves as a backdrop for a discussion on the role of signal processing techniques in the design of privacy-preserving media search systems.

Multiresolution graph signal processing via circulant structures

We use circulant structures to present a new framework for multiresolution analysis and processing of graph signals. Among the essential features of circulant graphs is that they accommodate fundamental signal processing operations, such as linear shift-invariant filtering, downsampling, upsampling, and reconstruction-features that offer substantial advantage. We design two-channel, critically-sampled, perfect-reconstruction, orthogonal lattice-filter structures to process signals defined on circulant graphs. To extend our reach to noncirculant graphs, we present a method to decompose a connected, undirected graph into a combination of circulant graphs. To evaluate our proposed framework, we offer examples of synthetic and real-world graph signal data and their multiscale decompositions.

Spline-Like Wavelet Filterbanks for Multiresolution Analysis of Graph-Structured Data

Multiresolution analysis is important for understanding graph signals, which represent graph-structured data. Wavelet filterbanks permit multiscale analysis and processing of graph signals-particularly, useful for harvesting large-scale data. Inspired by first-order spline wavelets in classical signal processing, we introduce two-channel (low-pass and high-pass) wavelet filterbanks for graph signals. This class of filterbanks boasts several useful properties, such as critical sampling, perfect reconstruction, and graph invariance. We consider an application in graph semi-supervised learning and propose a wavelet-regularized semi-supervised learning algorithm that is competitive for certain synthetic and real-world data.

Wavelet-regularized graph semi-supervised learning

Graph semi-supervised learning (GSSL) is a technique that uses a combination of labeled and unlabeled nodes on a graph to determine a classifier for new, incoming data. This problem can be analyzed through the lens of graph signal processing. In particular, the penalty functions used in the optimization formulation of standard GSSL algorithms can be interpreted as appropriately-defined filters in the Graph Fourier domain. We propose a wavelet-regularized semi-supervised learning algorithm using suitably-defined spline-like graph wavelets. These wavelets are critically-sampled, perfect-reconstruction basis representations, in contrast to much of the existing work proposing overcomplete representations. Critical sampling is essential for controlling the complexity in applications dealing with large scale datasets. We are also interested in understanding when wavelet-regularized approaches perform better than traditional Fourier-based regularizers. We compare the performance of our proposed spline-like, wavelet-regularized learning algorithm (as well as other existing graph wavelet designs) to some standard graph semi-supervised learning techniques on synthetic and real-world datasets.

Hiding the Rumor Source

Anonymous social media platforms, like Secret, Yik Yak, and Whisper, have emerged as important tools for sharing ideas without the fear of judgment. Such anonymous platforms are also important in nations under authoritarian rule, where freedom of expression and the personal safety of message that authors may depend on anonymity. Whether for fear of judgment or retribution, it is sometimes crucial to hide the identities of users who post sensitive messages. In this paper, we consider a global adversary who wishes to identify the author of a message; it observes either a snapshot of the spread of a message at a certain time or sampled timestamp metadata, or both. Recent advances in rumor source detection show that existing messaging protocols are vulnerable against such an adversary. We introduce a novel messaging protocol, which we call adaptive diffusion, and show that under the snapshot adversarial model, adaptive diffusion spreads content fast and achieves perfect obfuscation of the source when the underlying contact network is an infinite regular tree. That is, all users with the message are nearly equally likely to have been the origin of the message. When the contact network is an irregular tree, we characterize the probability of maximum likelihood detection by proving a concentration result over Galton–Watson trees. Experiments on a sampled Facebook network demonstrate that adaptive diffusion effectively hides the location of the source even when the graph is finite, is irregular, and has cycles.

Dandelion: Redesigning the Bitcoin Network for Anonymity

Cryptocurrencies are digital currencies that provide cryptographic verification of transactions. In recent years, they have transitioned from an academic research topic to a multi-billion dollar industry. Bitcoin is the best-known example of a cryptocurrency. Cryptocurrencies exhibit two key properties: egalitarianism and transparency. In this context, egalitarianism means that no single party wields disproportionate power over the networks operation. This diffusion of power is achieved by asking other network nodes (e.g., other Bitcoin users) to validate transactions, instead of the traditional method of using a centralized authority for this purpose. Moreover, all transactions and communications are managed over a fully-distributed, peer-to-peer (P2P) network. Cryptocurrencies are transparent in the sense that all transactions are verified and recorded with cryptographic integrity guarantees; this prevents fraudulent activity like double-spending of money. Transparency is achieved through a combination of clever cryptographic protocols and the publication of transactions in a ledger known as a blockchain. This blockchain serves as a public record of every financial transaction in the network. A property that Bitcoin does not provide is anonymity. Each user is identified in the network by a public, cryptographic key. If one were to link such a key to its owners human identity, the owners financial history could be partially learned from the public blockchain. In practice, it is possible to link public keys to identities through a number of channels, including the networking protocols on which Bitcoin is built. This is a massive privacy violation, and can be dangerous for deanonymized users.