Carmela Troncoso

Bogdan Kulynych, Juan Felipe Gomez, Georgios Kaissis et al.

Differential privacy (DP) is a widely used approach for mitigating privacy risks when training machine learning models on sensitive data. DP mechanisms add noise during training to limit the risk of information leakage. The scale of the added noise is critical, as it determines the trade-off between privacy and utility. The standard practice is to select the noise scale to satisfy a given privacy budget $\varepsilon$. This privacy budget is in turn interpreted in terms of operational attack risks, such as accuracy, sensitivity, and specificity of inference attacks aimed to recover information about the training data records. We show that first calibrating the noise scale to a privacy budget $\varepsilon$, and then translating ε to attack risk leads to overly conservative risk assessments and unnecessarily low utility. Instead, we propose methods to directly calibrate the noise scale to a desired attack risk level, bypassing the step of choosing $\varepsilon$. For a given notion of attack risk, our approach significantly decreases noise scale, leading to increased utility at the same level of privacy. We empirically demonstrate that calibrating noise to attack sensitivity/specificity, rather than $\varepsilon$, when training privacy-preserving ML models substantially improves model accuracy for the same risk level. Our work provides a principled and practical way to improve the utility of privacy-preserving ML without compromising on privacy. The code is available at https://github.com/Felipe-Gomez/riskcal

Dario Pasquini, Mathilde Raynal, Carmela Troncoso

In this work, we carry out the first, in-depth, privacy analysis of Decentralized Learning -- a collaborative machine learning framework aimed at addressing the main limitations of federated learning. We introduce a suite of novel attacks for both passive and active decentralized adversaries. We demonstrate that, contrary to what is claimed by decentralized learning proposers, decentralized learning does not offer any security advantage over federated learning. Rather, it increases the attack surface enabling any user in the system to perform privacy attacks such as gradient inversion, and even gain full control over honest users' local model. We also show that, given the state of the art in protections, privacy-preserving configurations of decentralized learning require fully connected networks, losing any practical advantage over the federated setup and therefore completely defeating the objective of the decentralized approach.

Bogdan Kulynych, Hsiang Hsu, Carmela Troncoso et al.

Mechanisms used in privacy-preserving machine learning often aim to guarantee differential privacy (DP) during model training. Practical DP-ensuring training methods use randomization when fitting model parameters to privacy-sensitive data (e.g., adding Gaussian noise to clipped gradients). We demonstrate that such randomization incurs predictive multiplicity: for a given input example, the output predicted by equally-private models depends on the randomness used in training. Thus, for a given input, the predicted output can vary drastically if a model is re-trained, even if the same training dataset is used. The predictive-multiplicity cost of DP training has not been studied, and is currently neither audited for nor communicated to model designers and stakeholders. We derive a bound on the number of re-trainings required to estimate predictive multiplicity reliably. We analyze--both theoretically and through extensive experiments--the predictive-multiplicity cost of three DP-ensuring algorithms: output perturbation, objective perturbation, and DP-SGD. We demonstrate that the degree of predictive multiplicity rises as the level of privacy increases, and is unevenly distributed across individuals and demographic groups in the data. Because randomness used to ensure DP during training explains predictions for some examples, our results highlight a fundamental challenge to the justifiability of decisions supported by differentially private models in high-stakes settings. We conclude that practitioners should audit the predictive multiplicity of their DP-ensuring algorithms before deploying them in applications of individual-level consequence.

Klim Kireev, Bogdan Kulynych, Carmela Troncoso

Many safety-critical applications of machine learning, such as fraud or abuse detection, use data in tabular domains. Adversarial examples can be particularly damaging for these applications. Yet, existing works on adversarial robustness primarily focus on machine-learning models in image and text domains. We argue that, due to the differences between tabular data and images or text, existing threat models are not suitable for tabular domains. These models do not capture that the costs of an attack could be more significant than imperceptibility, or that the adversary could assign different values to the utility obtained from deploying different adversarial examples. We demonstrate that, due to these differences, the attack and defense methods used for images and text cannot be directly applied to tabular settings. We address these issues by proposing new cost and utility-aware threat models that are tailored to the adversarial capabilities and constraints of attackers targeting tabular domains. We introduce a framework that enables us to design attack and defense mechanisms that result in models protected against cost and utility-aware adversaries, for example, adversaries constrained by a certain financial budget. We show that our approach is effective on three datasets corresponding to applications for which adversarial examples can have economic and social implications.

Dario Pasquini, Giuseppe Ateniese, Carmela Troncoso

We introduce the concept of "universal password model" -- a password model that, once pre-trained, can automatically adapt its guessing strategy based on the target system. To achieve this, the model does not need to access any plaintext passwords from the target credentials. Instead, it exploits users' auxiliary information, such as email addresses, as a proxy signal to predict the underlying password distribution. Specifically, the model uses deep learning to capture the correlation between the auxiliary data of a group of users (e.g., users of a web application) and their passwords. It then exploits those patterns to create a tailored password model for the target system at inference time. No further training steps, targeted data collection, or prior knowledge of the community's password distribution is required. Besides improving over current password strength estimation techniques and attacks, the model enables any end-user (e.g., system administrators) to autonomously generate tailored password models for their systems without the often unworkable requirements of collecting suitable training data and fitting the underlying machine learning model. Ultimately, our framework enables the democratization of well-calibrated password models to the community, addressing a major challenge in the deployment of password security solutions at scale.

Ana-Maria Cretu, Klim Kireev, Amro Abdalla et al.

We evaluate the effectiveness of filtering child images from training datasets of text-to-image models to prevent model misuse to create child sexual abuse material (CSAM). First, we capture the complexity of preventing CSAM generation using a game-based security definition. Second, we show that current detection methods cannot remove all children from a dataset. Third, using an ethical proxy for CSAM (a child wearing glasses), we show that even when only a small percentage of child images are left in the training dataset after filtering, there exist prompting strategies that generate a child wearing glasses using only a few more queries than when the model is trained on the unfiltered data. Fine-tuning the filtered model on child images further reduces the additional query overhead. We also show that re-introducing a concept is possible via fine-tuning even if filtering is perfect. Our results show that current child filtering methods offer limited protection to closed-weight models and no protection to open-weight models, while reducing the generality of the model by hindering the generation of child-related concepts or changing their representation. We conclude by outlining challenges in conducting evaluations that establish robust evidence on the impact of concept filtering defenses for CSAM.

Mathilde Raynal, Dario Pasquini, Carmela Troncoso

Decentralized Learning (DL) is a peer--to--peer learning approach that allows a group of users to jointly train a machine learning model. To ensure correctness, DL should be robust, i.e., Byzantine users must not be able to tamper with the result of the collaboration. In this paper, we introduce two \textit{new} attacks against DL where a Byzantine user can: make the network converge to an arbitrary model of their choice, and exclude an arbitrary user from the learning process. We demonstrate our attacks' efficiency against Self--Centered Clipping, the state--of--the--art robust DL protocol. Finally, we show that the capabilities decentralization grants to Byzantine users result in decentralized learning \emph{always} providing less robustness than federated learning.

Klim Kireev, Maksym Andriushchenko, Carmela Troncoso et al.

Research on adversarial robustness is primarily focused on image and text data. Yet, many scenarios in which lack of robustness can result in serious risks, such as fraud detection, medical diagnosis, or recommender systems often do not rely on images or text but instead on tabular data. Adversarial robustness in tabular data poses two serious challenges. First, tabular datasets often contain categorical features, and therefore cannot be tackled directly with existing optimization procedures. Second, in the tabular domain, algorithms that are not based on deep networks are widely used and offer great performance, but algorithms to enhance robustness are tailored to neural networks (e.g. adversarial training). In this paper, we tackle both challenges. We present a method that allows us to train adversarially robust deep networks for tabular data and to transfer this robustness to other classifiers via universal robust embeddings tailored to categorical data. These embeddings, created using a bilevel alternating minimization framework, can be transferred to boosted trees or random forests making them robust without the need for adversarial training while preserving their high accuracy on tabular data. We show that our methods outperform existing techniques within a practical threat model suitable for tabular data.

Bogdan Kulynych, Theresa Stadler, Jean Louis Raisaro et al.

Recent advances in generative modelling have led many to see synthetic data as the go-to solution for a range of problems around data access, scarcity, and under-representation. In this paper, we study three prominent use cases: (1) Sharing synthetic data as a proxy for proprietary datasets to enable statistical analyses while protecting privacy, (2) Augmenting machine learning training sets with synthetic data to improve model performance, and (3) Augmenting datasets with synthetic data to reduce variance in statistical estimation. For each use case, we formalise the problem setting and study, through formal analysis and case studies, under which conditions synthetic data can achieve its intended objectives. We identify fundamental and practical limits that constrain when synthetic data can serve as an effective solution for a particular problem. Our analysis reveals that due to these limits many existing or envisioned use cases of synthetic data are a poor problem fit. Our formalisations and classification of synthetic data use cases enable decision makers to assess whether synthetic data is a suitable approach for their specific data availability problem.

Wouter Lueks, Bogdan Kulynych, Jules Fasquelle et al.

Zero-knowledge proofs are an essential building block in many privacy-preserving systems. However, implementing these proofs is tedious and error-prone. In this paper, we present zksk, a well-documented Python library for defining and computing sigma protocols: the most popular class of zero-knowledge proofs. In zksk, proofs compose: programmers can convert smaller proofs into building blocks that then can be combined into bigger proofs. zksk features a modern Python-based domain-specific language. This makes possible to define proofs without learning a new custom language, and to benefit from the rich Python syntax and ecosystem. The library is available at https://github.com/spring-epfl/zksk

Bogdan Kulynych, Mohammad Yaghini, Giovanni Cherubin et al.

A membership inference attack (MIA) against a machine-learning model enables an attacker to determine whether a given data record was part of the model's training data or not. In this paper, we provide an in-depth study of the phenomenon of disparate vulnerability against MIAs: unequal success rate of MIAs against different population subgroups. We first establish necessary and sufficient conditions for MIAs to be prevented, both on average and for population subgroups, using a notion of distributional generalization. Second, we derive connections of disparate vulnerability to algorithmic fairness and to differential privacy. We show that fairness can only prevent disparate vulnerability against limited classes of adversaries. Differential privacy bounds disparate vulnerability but can significantly reduce the accuracy of the model. We show that estimating disparate vulnerability to MIAs by naïvely applying existing attacks can lead to overestimation. We then establish which attacks are suitable for estimating disparate vulnerability, and provide a statistical framework for doing so reliably. We conduct experiments on synthetic and real-world data finding statistically significant evidence of disparate vulnerability in realistic settings. The code is available at https://github.com/spring-epfl/disparate-vulnerability

Dario Pasquini, Martin Strohmeier, Carmela Troncoso

We introduce a new family of prompt injection attacks, termed Neural Exec. Unlike known attacks that rely on handcrafted strings (e.g., "Ignore previous instructions and..."), we show that it is possible to conceptualize the creation of execution triggers as a differentiable search problem and use learning-based methods to autonomously generate them. Our results demonstrate that a motivated adversary can forge triggers that are not only drastically more effective than current handcrafted ones but also exhibit inherent flexibility in shape, properties, and functionality. In this direction, we show that an attacker can design and generate Neural Execs capable of persisting through multi-stage preprocessing pipelines, such as in the case of Retrieval-Augmented Generation (RAG)-based applications. More critically, our findings show that attackers can produce triggers that deviate markedly in form and shape from any known attack, sidestepping existing blacklist-based detection and sanitation approaches.

Saiid El Hajj Chehade, Sandra Siby, Carmela Troncoso

Privacy-enhancing blocking tools based on filter-list rules tend to break legitimate functionality. Filter-list maintainers could benefit from automated breakage detection tools that allow them to proactively fix problematic rules before deploying them to millions of users. We introduce SINBAD, an automated breakage detector that improves the accuracy over the state of the art by 20%, and is the first to detect dynamic breakage and breakage caused by style-oriented filter rules. The success of SINBAD is rooted in three innovations: (1) the use of user-reported breakage issues in forums that enable the creation of a high-quality dataset for training in which only breakage that users perceive as an issue is included; (2) the use of 'web saliency' to automatically identify user-relevant regions of a website on which to prioritize automated interactions aimed at triggering breakage; and (3) the analysis of webpages via subtrees which enables fine-grained identification of problematic filter rules.

Mathilde Raynal, Carmela Troncoso

Collaborative Machine Learning (CML) allows participants to jointly train a machine learning model while keeping their training data private. In many scenarios where CML is seen as the solution to privacy issues, such as health-related applications, safety is also a primary concern. To ensure that CML processes produce models that output correct and reliable decisions \emph{even in the presence of potentially untrusted participants}, researchers propose to use \textit{robust aggregators} to filter out malicious contributions that negatively influence the training process. In this work, we formalize the two prevalent forms of robust aggregators in the literature. We then show that neither can provide the intended protection: either they use distance-based metrics that cannot reliably identify malicious inputs to training; or use metrics based on the behavior of the loss function which create a conflict with the ability of CML participants to learn, i.e., they cannot eliminate the risk of compromise without preventing learning.

Theresa Stadler, Bogdan Kulynych, Michael C. Gastpar et al.

The promise of least-privilege learning -- to find feature representations that are useful for a learning task but prevent inference of any sensitive information unrelated to this task -- is highly appealing. However, so far this concept has only been stated informally. It thus remains an open question whether and how we can achieve this goal. In this work, we provide the first formalisation of the least-privilege principle for machine learning and characterise its feasibility. We prove that there is a fundamental trade-off between a representation's utility for a given task and its leakage beyond the intended task: it is not possible to learn representations that have high utility for the intended task but, at the same time prevent inference of any attribute other than the task label itself. This trade-off holds under realistic assumptions on the data distribution and regardless of the technique used to learn the feature mappings that produce these representations. We empirically validate this result for a wide range of learning techniques, model architectures, and datasets.

Klim Kireev, Ana-Maria Creţu, Raphael Meier et al.

Platforms and the law regulate digital content depicting minors (defined as individuals under 18 years of age) differently from other types of content. Given the sheer amount of content that needs to be assessed, machine learning-based automation tools are commonly used to detect content depicting minors. To our knowledge, no dataset or benchmark currently exists for detecting these identification methods in a multi-modal environment. To fill this gap, we release the Image-Caption Children in the Wild Dataset (ICCWD), an image-caption dataset aimed at benchmarking tools that detect depictions of minors. Our dataset is richer than previous child image datasets, containing images of children in a variety of contexts, including fictional depictions and partially visible bodies. ICCWD contains 10,000 image-caption pairs manually labeled to indicate the presence or absence of a child in the image. To demonstrate the possible utility of our dataset, we use it to benchmark three different detectors, including a commercial age estimation system applied to images. Our results suggest that child detection is a challenging task, with the best method achieving a 75.3% true positive rate. We hope the release of our dataset will aid in the design of better minor detection methods in a wide range of scenarios.

Hal Abelson, Ross Anderson, Steven M. Bellovin et al.

Our increasing reliance on digital technology for personal, economic, and government affairs has made it essential to secure the communications and devices of private citizens, businesses, and governments. This has led to pervasive use of cryptography across society. Despite its evident advantages, law enforcement and national security agencies have argued that the spread of cryptography has hindered access to evidence and intelligence. Some in industry and government now advocate a new technology to access targeted data: client-side scanning (CSS). Instead of weakening encryption or providing law enforcement with backdoor keys to decrypt communications, CSS would enable on-device analysis of data in the clear. If targeted information were detected, its existence and, potentially, its source, would be revealed to the agencies; otherwise, little or no information would leave the client device. Its proponents claim that CSS is a solution to the encryption versus public safety debate: it offers privacy -- in the sense of unimpeded end-to-end encryption -- and the ability to successfully investigate serious crime. In this report, we argue that CSS neither guarantees efficacious crime prevention nor prevents surveillance. Indeed, the effect is the opposite. CSS by its nature creates serious security and privacy risks for all society while the assistance it can provide for law enforcement is at best problematic. There are multiple ways in which client-side scanning can fail, can be evaded, and can be abused.

Sandra Siby, Umar Iqbal, Steven Englehardt et al.

Millions of web users directly depend on ad and tracker blocking tools to protect their privacy. However, existing ad and tracker blockers fall short because of their reliance on trivially susceptible advertising and tracking content. In this paper, we first demonstrate that the state-of-the-art machine learning based ad and tracker blockers, such as AdGraph, are susceptible to adversarial evasions deployed in real-world. Second, we introduce WebGraph, the first graph-based machine learning blocker that detects ads and trackers based on their action rather than their content. By building features around the actions that are fundamental to advertising and tracking - storing an identifier in the browser, or sharing an identifier with another tracker - WebGraph performs nearly as well as prior approaches, but is significantly more robust to adversarial evasions. In particular, we show that WebGraph achieves comparable accuracy to AdGraph, while significantly decreasing the success rate of an adversary from near-perfect under AdGraph to around 8% under WebGraph. Finally, we show that WebGraph remains robust to a more sophisticated adversary that uses evasion techniques beyond those currently deployed on the web.

Theresa Stadler, Wouter Lueks, Katharina Kohls et al.

In this document, we analyse the potential harms a large-scale deployment of the Luca system might cause to individuals, venues, and communities. The Luca system is a digital presence tracing system designed to provide health departments with the contact information necessary to alert individuals who have visited a location at the same time as a SARS-CoV-2-positive person. Multiple regional health departments in Germany have announced their plans to deploy the Luca system for the purpose of presence tracing. The system's developers suggest its use across various types of venues: from bars and restaurants to public and private events, such religious or political gatherings, weddings, and birthday parties. Recently, an extension to include schools and other educational facilities was discussed in public. Our analysis of the potential harms of the system is based on the publicly available Luca Security Concept which describes the system's security architecture and its planned protection mechanisms. The Security Concept furthermore provides a set of claims about the system's security and privacy properties. Besides an analysis of harms, our analysis includes a validation of these claims.

Theresa Stadler, Bristena Oprisanu, Carmela Troncoso

Synthetic data has been advertised as a silver-bullet solution to privacy-preserving data publishing that addresses the shortcomings of traditional anonymisation techniques. The promise is that synthetic data drawn from generative models preserves the statistical properties of the original dataset but, at the same time, provides perfect protection against privacy attacks. In this work, we present the first quantitative evaluation of the privacy gain of synthetic data publishing and compare it to that of previous anonymisation techniques. Our evaluation of a wide range of state-of-the-art generative models demonstrates that synthetic data either does not prevent inference attacks or does not retain data utility. In other words, we empirically show that synthetic data does not provide a better tradeoff between privacy and utility than traditional anonymisation techniques. Furthermore, in contrast to traditional anonymisation, the privacy-utility tradeoff of synthetic data publishing is hard to predict. Because it is impossible to predict what signals a synthetic dataset will preserve and what information will be lost, synthetic data leads to a highly variable privacy gain and unpredictable utility loss. In summary, we find that synthetic data is far from the holy grail of privacy-preserving data publishing.

Konstantinos Chatzikokolakis, Giovanni Cherubin, Catuscia Palamidessi et al.

Security system designers favor worst-case security metrics, such as those derived from differential privacy (DP), due to the strong guarantees they provide. On the downside, these guarantees result in a high penalty on the system's performance. In this paper, we study Bayes security, a security metric inspired by the cryptographic advantage. Similarly to DP, Bayes security i) is independent of an adversary's prior knowledge, ii) it captures the worst-case scenario for the two most vulnerable secrets (e.g., data records); and iii) it is easy to compose, facilitating security analyses. Additionally, Bayes security iv) can be consistently estimated in a black-box manner, contrary to DP, which is useful when a formal analysis is not feasible; and v) provides a better utility-security trade-off in high-security regimes because it quantifies the risk for a specific threat model as opposed to threat-agnostic metrics such as DP. We formulate a theory around Bayes security, and we provide a thorough comparison with respect to well-known metrics, identifying the scenarios where Bayes Security is advantageous for designers.

Blagovesta Kostova, Seda Gürses, Carmela Troncoso

Current day software development relies heavily on the use of service architectures and on agile iterative development methods to design, implement, and deploy systems. These practices result in systems made up of multiple services that introduce new data flows and evolving designs that escape the control of a single designer. Academic privacy engineering literature typically abstracts away such conditions of software production in order to achieve generalizable results. Yet, through a systematic study of the literature, we show that proposed solutions inevitably make assumptions about software architectures, development methods and scope of designer control that are misaligned with current practices. These misalignments are likely to pose an obstacle to operationalizing privacy engineering solutions in the wild. Specifically, we identify important limitations in the approaches that researchers take to design and evaluate privacy enhancing technologies which ripple to proposals for privacy engineering methodologies. Based on our analysis, we delineate research and actions needed to re-align research with practice, changes that serve a precondition for the operationalization of academic privacy results in common software engineering practices.

Kasra EdalatNejad, Wouter Lueks, Julien Pierre Martin et al.

Investigative journalists collect large numbers of digital documents during their investigations. These documents can greatly benefit other journalists' work. However, many of these documents contain sensitive information. Hence, possessing such documents can endanger reporters, their stories, and their sources. Consequently, many documents are used only for single, local, investigations. We present DatashareNetwork, a decentralized and privacy-preserving search system that enables journalists worldwide to find documents via a dedicated network of peers. DatashareNetwork combines well-known anonymous authentication mechanisms and anonymous communication primitives, a novel asynchronous messaging system, and a novel multi-set private set intersection protocol (MS-PSI) into a *decentralized peer-to-peer private document search engine*. We prove that DatashareNetwork is secure; and show using a prototype implementation that it scales to thousands of users and millions of documents.

Carmela Troncoso, Mathias Payer, Jean-Pierre Hubaux et al.

This document describes and analyzes a system for secure and privacy-preserving proximity tracing at large scale. This system, referred to as DP3T, provides a technological foundation to help slow the spread of SARS-CoV-2 by simplifying and accelerating the process of notifying people who might have been exposed to the virus so that they can take appropriate measures to break its transmission chain. The system aims to minimise privacy and security risks for individuals and communities and guarantee the highest level of data protection. The goal of our proximity tracing system is to determine who has been in close physical proximity to a COVID-19 positive person and thus exposed to the virus, without revealing the contact's identity or where the contact occurred. To achieve this goal, users run a smartphone app that continually broadcasts an ephemeral, pseudo-random ID representing the user's phone and also records the pseudo-random IDs observed from smartphones in close proximity. When a patient is diagnosed with COVID-19, she can upload pseudo-random IDs previously broadcast from her phone to a central server. Prior to the upload, all data remains exclusively on the user's phone. Other users' apps can use data from the server to locally estimate whether the device's owner was exposed to the virus through close-range physical proximity to a COVID-19 positive person who has uploaded their data. In case the app detects a high risk, it will inform the user.

Wouter Lueks, Iñigo Querejeta-Azurmendi, Carmela Troncoso

The strongest threat model for voting systems considers coercion resistance: protection against coercers that force voters to modify their votes, or to abstain. Existing remote voting systems either do not provide this property; require an expensive tallying phase; or burden users with the need to store cryptographic key material and with the responsibility to deceive their coercers. We propose VoteAgain, a scalable voting scheme that relies on the revoting paradigm to provide coercion resistance. VoteAgain uses a novel deterministic ballot padding mechanism to ensure that coercers cannot see whether a vote has been replaced. This mechanism ensures tallies take quasilinear time, making VoteAgain the first revoting scheme that can handle elections with millions of voters. We prove that VoteAgain provides ballot privacy, coercion resistance, and verifiability; and we demonstrate its scalability using a prototype implementation of all cryptographic primitives.

Simon Oya, Fernando Pérez-González, Carmela Troncoso

In this work, we address the problem of designing delay-based anonymous communication systems. We consider a timed mix where an eavesdropper wants to learn the communication pattern of the users, and study how the mix must delay the messages so as to increase the adversary's estimation error. We show the connection between this problem and a MIMO system where we want to design the coloring filter that worsens the adversary's estimation of the MIMO channel matrix. We obtain theoretical solutions for the optimal filter against short-term and long-term adversaries, evaluate them with experiments, and show how some properties of filters can be used in the implementation of timed mixes. This opens the door to the application of previously known filter design techniques to anonymous communication systems.

Simon Oya, Carmela Troncoso, Fernando Pérez-González

High-latency anonymous communication systems prevent passive eavesdroppers from inferring communicating partners with certainty. However, disclosure attacks allow an adversary to recover users' behavioral profiles when communications are persistent. Understanding how the system parameters affect the privacy of the users against such attacks is crucial. Earlier work in the area analyzes the performance of disclosure attacks in controlled scenarios, where a certain model about the users' behavior is assumed. In this paper, we analyze the profiling accuracy of one of the most efficient disclosure attack, the least squares disclosure attack, in realistic scenarios. We generate real traffic observations from datasets of different nature and find that the models considered in previous work do not fit this realistic behavior. We relax previous hypotheses on the behavior of the users and extend previous performance analyses, validating our results with real data and providing new insights into the parameters that affect the protection of the users in the real world.

Fernando Pérez-González, Carmela Troncoso, Simon Oya

Mixes, relaying routers that hide the relation between incoming and outgoing messages, are the main building block of high-latency anonymous communication networks. A number of so-called disclosure attacks have been proposed to effectively de-anonymize traffic sent through these channels. Yet, the dependence of their success on the system parameters is not well-understood. We propose the Least Squares Disclosure Attack (LSDA), in which user profiles are estimated by solving a least squares problem. We show that LSDA is not only suitable for the analysis of threshold mixes, but can be easily extended to attack pool mixes. Furthermore, contrary to previous heuristic-based attacks, our approach allows us to analytically derive expressions that characterize the profiling error of LSDA with respect to the system parameters. We empirically demonstrate that LSDA recovers users' profiles with greater accuracy than its statistical predecessors and verify that our analysis closely predicts actual performance.

Simon Oya, Carmela Troncoso, Fernando Pérez-González

Disclosure attacks aim at revealing communication patterns in anonymous communication systems, such as conversation partners or frequency. In this paper, we propose a framework to compare between the members of the statistical disclosure attack family. We compare different variants of the Statistical Disclosure Attack (SDA) in the literature, together with two new methods; as well as show their relation with the Least Squares Disclosure Attack (LSDA). We empirically explore the performance of the attacks with respect to the different parameters of the system. Our experiments show that i) our proposals considerably improve the state-of-the-art SDA and ii) confirm that LSDA outperforms the SDA family when the adversary has enough observations of the system.

Sandra Siby, Marc Juarez, Claudia Diaz et al.

Virtually every connection to an Internet service is preceded by a DNS lookup which is performed without any traffic-level protection, thus enabling manipulation, redirection, surveillance, and censorship. To address these issues, large organizations such as Google and Cloudflare are deploying recently standardized protocols that encrypt DNS traffic between end users and recursive resolvers such as DNS-over-TLS (DoT) and DNS-over-HTTPS (DoH). In this paper, we examine whether encrypting DNS traffic can protect users from traffic analysis-based monitoring and censoring. We propose a novel feature set to perform the attacks, as those used to attack HTTPS or Tor traffic are not suitable for DNS' characteristics. We show that traffic analysis enables the identification of domains with high accuracy in closed and open world settings, using 124 times less data than attacks on HTTPS flows. We find that factors such as location, resolver, platform, or client do mitigate the attacks performance but they are far from completely stopping them. Our results indicate that DNS-based censorship is still possible on encrypted DNS traffic. In fact, we demonstrate that the standardized padding schemes are not effective. Yet, Tor -- which does not effectively mitigate traffic analysis attacks on web traffic -- is a good defense against DoH traffic analysis.

Apostolos Pyrgelis, Carmela Troncoso, Emiliano De Cristofaro

While location data is extremely valuable for various applications, disclosing it prompts serious threats to individuals' privacy. To limit such concerns, organizations often provide analysts with aggregate time-series that indicate, e.g., how many people are in a location at a time interval, rather than raw individual traces. In this paper, we perform a measurement study to understand Membership Inference Attacks (MIAs) on aggregate location time-series, where an adversary tries to infer whether a specific user contributed to the aggregates. We find that the volume of contributed data, as well as the regularity and particularity of users' mobility patterns, play a crucial role in the attack's success. We experiment with a wide range of defenses based on generalization, hiding, and perturbation, and evaluate their ability to thwart the attack vis-a-vis the utility loss they introduce for various mobility analytics tasks. Our results show that some defenses fail across the board, while others work for specific tasks on aggregate location time-series. For instance, suppressing small counts can be used for ranking hotspots, data generalization for forecasting traffic, hotspot discovery, and map inference, while sampling is effective for location labeling and anomaly detection when the dataset is sparse. Differentially private techniques provide reasonable accuracy only in very specific settings, e.g., discovering hotspots and forecasting their traffic, and more so when using weaker privacy notions like crowd-blending privacy. Overall, our measurements show that there does not exist a unique generic defense that can preserve the utility of the analytics for arbitrary applications, and provide useful insights regarding the disclosure of sanitized aggregate location time-series.

Spyros Boukoros, Mathias Humbert, Stefan Katzenbeisser et al.

Crowdsourcing enables application developers to benefit from large and diverse datasets at a low cost. Specifically, mobile crowdsourcing (MCS) leverages users' devices as sensors to perform geo-located data collection. The collection of geolocated data raises serious privacy concerns for users. Yet, despite the large research body on location privacy-preserving mechanisms (LPPMs), MCS developers implement little to no protection for data collection or publication. To understand this mismatch, we study the performance of existing LPPMs on publicly available data from two mobile crowdsourcing projects. Our results show that well-established defenses are either not applicable or offer little protection in the MCS setting. Additionally, they have a much stronger impact on applications' utility than foreseen in the literature. This is because existing LPPMs, designed with location-based services (LBSs) in mind, are optimized for utility functions based on users' locations, while MCS utility functions depend on the values (e.g., measurements) associated with those locations. We finally outline possible research avenues to facilitate the development of new location privacy solutions that fit the needs of MCS so that the increasing number of such applications do not jeopardize their users' privacy.

Rebekah Overdorf, Bogdan Kulynych, Ero Balsa et al.

In addition to their benefits, optimization systems can have negative economic, moral, social, and political effects on populations as well as their environments. Frameworks like fairness have been proposed to aid service providers in addressing subsequent bias and discrimination during data collection and algorithm design. However, recent reports of neglect, unresponsiveness, and malevolence cast doubt on whether service providers can effectively implement fairness solutions. These reports invite us to revisit assumptions made about the service providers in fairness solutions. Namely, that service providers have (i) the incentives or (ii) the means to mitigate optimization externalities. Moreover, the environmental impact of these systems suggests that we need (iii) novel frameworks that consider systems other than algorithmic decision-making and recommender systems, and (iv) solutions that go beyond removing related algorithmic biases. Going forward, we propose Protective Optimization Technologies that enable optimization subjects to defend against negative consequences of optimization systems.

Bogdan Kulynych, Jamie Hayes, Nikita Samarin et al.

Machine-learning models for security-critical applications such as bot, malware, or spam detection, operate in constrained discrete domains. These applications would benefit from having provable guarantees against adversarial examples. The existing literature on provable adversarial robustness of models, however, exclusively focuses on robustness to gradient-based attacks in domains such as images. These attacks model the adversarial cost, e.g., amount of distortion applied to an image, as a $p$-norm. We argue that this approach is not well-suited to model adversarial costs in constrained domains where not all examples are feasible. We introduce a graphical framework that (1) generalizes existing attacks in discrete domains, (2) can accommodate complex cost functions beyond $p$-norms, including financial cost incurred when attacking a classifier, and (3) efficiently produces valid adversarial examples with guarantees of minimal adversarial cost. These guarantees directly translate into a notion of adversarial robustness that takes into account domain constraints and the adversary's capabilities. We show how our framework can be used to evaluate security by crafting adversarial examples that evade a Twitter-bot detection classifier with provably minimal number of changes; and to build privacy defenses by crafting adversarial examples that evade a privacy-invasive website-fingerprinting classifier.

Simon Oya, Carmela Troncoso, Fernando Pérez-González

Location Privacy-Preserving Mechanisms (LPPMs) in the literature largely consider that users' data available for training wholly characterizes their mobility patterns. Thus, they hardwire this information in their designs and evaluate their privacy properties with these same data. In this paper, we aim to understand the impact of this decision on the level of privacy these LPPMs may offer in real life when the users' mobility data may be different from the data used in the design phase. Our results show that, in many cases, training data does not capture users' behavior accurately and, thus, the level of privacy provided by the LPPM is often overestimated. To address this gap between theory and practice, we propose to use blank-slate models for LPPM design. Contrary to the hardwired approach, that assumes known users' behavior, blank-slate models learn the users' behavior from the queries to the service provider. We leverage this blank-slate approach to develop a new family of LPPMs, that we call Profile Estimation-Based LPPMs. Using real data, we empirically show that our proposal outperforms optimal state-of-the-art mechanisms designed on sporadic hardwired models. On non-sporadic location privacy scenarios, our method is only better if the usage of the location privacy service is not continuous. It is our hope that eliminating the need to bootstrap the mechanisms with training data and ensuring that the mechanisms are lightweight and easy to compute help fostering the integration of location privacy protections in deployed systems.

Wouter Lueks, Brinda Hampiholi, Greg Alpár et al.

Users' devices, e.g., smartphones or laptops, are typically incapable of securely storing and processing cryptographic keys. We present Tandem, a novel set of protocols for securing cryptographic keys with support from a central server. Tandem uses one-time-use key-share tokens to preserve users' privacy with respect to a malicious central server. Additionally, Tandem enables users to block their keys if they lose their device, and it enables the server to limit how often an adversary can use an unblocked key. We prove Tandem's security and privacy properties, apply Tandem to attribute-based credentials, and implement a Tandem proof of concept to show that it causes little overhead.

Bogdan Kulynych, Rebekah Overdorf, Carmela Troncoso et al.

Algorithmic fairness aims to address the economic, moral, social, and political impact that digital systems have on populations through solutions that can be applied by service providers. Fairness frameworks do so, in part, by mapping these problems to a narrow definition and assuming the service providers can be trusted to deploy countermeasures. Not surprisingly, these decisions limit fairness frameworks' ability to capture a variety of harms caused by systems. We characterize fairness limitations using concepts from requirements engineering and from social sciences. We show that the focus on algorithms' inputs and outputs misses harms that arise from systems interacting with the world; that the focus on bias and discrimination omits broader harms on populations and their environments; and that relying on service providers excludes scenarios where they are not cooperative or intentionally adversarial. We propose Protective Optimization Technologies (POTs). POTs provide means for affected parties to address the negative impacts of systems in the environment, expanding avenues for political contestation. POTs intervene from outside the system, do not require service providers to cooperate, and can serve to correct, shift, or expose harms that systems impose on populations and their environments. We illustrate the potential and limitations of POTs in two case studies: countering road congestion caused by traffic-beating applications, and recalibrating credit scoring for loan applicants.

Rebekah Overdorf, Carmela Troncoso, Rachel Greenstadt et al.

Underground forums where users discuss, buy, and sell illicit services and goods facilitate a better understanding of the economy and organization of cybercriminals. Prior work has shown that in particular private interactions provide a wealth of information about the cybercriminal ecosystem. Yet, those messages are seldom available to analysts, except when there is a leak. To address this problem we propose a supervised machine learning based method able to predict which public \threads will generate private messages, after a partial leak of such messages has occurred. To the best of our knowledge, we are the first to develop a solution to overcome the barrier posed by limited to no information on private activity for underground forum analysis. Additionally, we propose an automate method for labeling posts, significantly reducing the cost of our approach in the presence of real unlabeled data. This method can be tuned to focus on the likelihood of users receiving private messages, or \threads triggering private interactions. We evaluate the performance of our methods using data from three real forum leaks. Our results show that public information can indeed be used to predict private activity, although prediction models do not transfer well between forums. We also find that neither the length of the leak period nor the time between the leak and the prediction have significant impact on our technique's performance, and that NLP features dominate the prediction power.

Chen Chen, Daniele E. Asoni, Adrian Perrig et al.

Modern low-latency anonymity systems, no matter whether constructed as an overlay or implemented at the network layer, offer limited security guarantees against traffic analysis. On the other hand, high-latency anonymity systems offer strong security guarantees at the cost of computational overhead and long delays, which are excessive for interactive applications. We propose TARANET, an anonymity system that implements protection against traffic analysis at the network layer, and limits the incurred latency and overhead. In TARANET's setup phase, traffic analysis is thwarted by mixing. In the data transmission phase, end hosts and ASes coordinate to shape traffic into constant-rate transmission using packet splitting. Our prototype implementation shows that TARANET can forward anonymous traffic at over 50~Gbps using commodity hardware.

Bogdan Kulynych, Carmela Troncoso

Understanding the influence of features in machine learning is crucial to interpreting models and selecting the best features for classification. In this work we propose the use of principles from coalitional game theory to reason about importance of features. In particular, we propose the use of the Banzhaf power index as a measure of influence of features on the outcome of a classifier. We show that features having Banzhaf power index of zero can be losslessly pruned without damage to classifier accuracy. Computing the power indices does not require having access to data samples. However, if samples are available, the indices can be empirically estimated. We compute Banzhaf power indices for a neural network classifier on real-life data, and compare the results with gradient-based feature saliency, and coefficients of a logistic regression model with $L_1$ regularization.

Simon Oya, Carmela Troncoso, Fernando Pérez-González

Since its proposal in 2013, geo-indistinguishability has been consolidated as a formal notion of location privacy, generating a rich body of literature building on this idea. A problem with most of these follow-up works is that they blindly rely on geo-indistinguishability to provide location privacy, ignoring the numerical interpretation of this privacy guarantee. In this paper, we provide an alternative formulation of geo-indistinguishability as an adversary error, and use it to show that the privacy vs.~utility trade-off that can be obtained is not as appealing as implied by the literature. We also show that although geo-indistinguishability guarantees a lower bound on the adversary's error, this comes at the cost of achieving poorer performance than other noise generation mechanisms in terms of average error, and enabling the possibility of exposing obfuscated locations that are useless from the quality of service point of view.

Apostolos Pyrgelis, Carmela Troncoso, Emiliano De Cristofaro

Aggregate location data is often used to support smart services and applications, e.g., generating live traffic maps or predicting visits to businesses. In this paper, we present the first study on the feasibility of membership inference attacks on aggregate location time-series. We introduce a game-based definition of the adversarial task, and cast it as a classification problem where machine learning can be used to distinguish whether or not a target user is part of the aggregates. We empirically evaluate the power of these attacks on both raw and differentially private aggregates using two mobility datasets. We find that membership inference is a serious privacy threat, and show how its effectiveness depends on the adversary's prior knowledge, the characteristics of the underlying location data, as well as the number of users and the timeframe on which aggregation is performed. Although differentially private mechanisms can indeed reduce the extent of the attacks, they also yield a significant loss in utility. Moreover, a strategic adversary mimicking the behavior of the defense mechanism can greatly limit the protection they provide. Overall, our work presents a novel methodology geared to evaluate membership inference on aggregate location data in real-world settings and can be used by providers to assess the quality of privacy protection before data release or by regulators to detect violations.

Bogdan Kulynych, Wouter Lueks, Marios Isaakidis et al.

The social demand for email end-to-end encryption is barely supported by mainstream service providers. Autocrypt is a new community-driven open specification for e-mail encryption that attempts to respond to this demand. In Autocrypt the encryption keys are attached directly to messages, and thus the encryption can be implemented by email clients without any collaboration of the providers. The decentralized nature of this in-band key distribution, however, makes it prone to man-in-the-middle attacks and can leak the social graph of users. To address this problem we introduce ClaimChain, a cryptographic construction for privacy-preserving authentication of public keys. Users store claims about their identities and keys, as well as their beliefs about others, in ClaimChains. These chains form authenticated decentralized repositories that enable users to prove the authenticity of both their keys and the keys of their contacts. ClaimChains are encrypted, and therefore protect the stored information, such as keys and contact identities, from prying eyes. At the same time, ClaimChain implements mechanisms to provide strong non-equivocation properties, discouraging malicious actors from distributing conflicting or inauthentic claims. We implemented ClaimChain and we show that it offers reasonable performance, low overhead, and authenticity guarantees.

Simon Oya, Carmela Troncoso, Fernando Pérez-González

In the last years we have witnessed the appearance of a variety of strategies to design optimal location privacy-preserving mechanisms, in terms of maximizing the adversary's expected error with respect to the users' whereabouts. In this work, we take a closer look at the defenses created by these strategies and show that, even though they are indeed optimal in terms of adversary's correctness, not all of them offer the same protection when looking at other dimensions of privacy. To avoid "bad" choices, we argue that the search for optimal mechanisms must be guided by complementary criteria. We provide two example auxiliary metrics that help in this regard: the conditional entropy, that captures an information-theoretic aspect of the problem; and the worst-case quality loss, that ensures that the output of the mechanism always provides a minimum utility to the users. We describe a new mechanism that maximizes the conditional entropy and is optimal in terms of average adversary error, and compare its performance with previously proposed optimal mechanisms using two real datasets. Our empirical results confirm that no mechanism fares well on every privacy criteria simultaneously, making apparent the need for considering multiple privacy dimensions to have a good understanding of the privacy protection a mechanism provides.

Carmela Troncoso, Marios Isaakidis, George Danezis et al.

Decentralized systems are a subset of distributed systems where multiple authorities control different components and no authority is fully trusted by all. This implies that any component in a decentralized system is potentially adversarial. We revise fifteen years of research on decentralization and privacy, and provide an overview of key systems, as well as key insights for designers of future systems. We show that decentralized designs can enhance privacy, integrity, and availability but also require careful trade-offs in terms of system complexity, properties provided, and degree of decentralization. These trade-offs need to be understood and navigated by designers. We argue that a combination of insights from cryptography, distributed systems, and mechanism design, aligned with the development of adequate incentives, are necessary to build scalable and successful privacy-preserving decentralized systems.

Apostolos Pyrgelis, Carmela Troncoso, Emiliano De Cristofaro

Information about people's movements and the locations they visit enables an increasing number of mobility analytics applications, e.g., in the context of urban and transportation planning, In this setting, rather than collecting or sharing raw data, entities often use aggregation as a privacy protection mechanism, aiming to hide individual users' location traces. Furthermore, to bound information leakage from the aggregates, they can perturb the input of the aggregation or its output to ensure that these are differentially private. In this paper, we set to evaluate the impact of releasing aggregate location time-series on the privacy of individuals contributing to the aggregation. We introduce a framework allowing us to reason about privacy against an adversary attempting to predict users' locations or recover their mobility patterns. We formalize these attacks as inference problems, and discuss a few strategies to model the adversary's prior knowledge based on the information she may have access to. We then use the framework to quantify the privacy loss stemming from aggregate location data, with and without the protection of differential privacy, using two real-world mobility datasets. We find that aggregates do leak information about individuals' punctual locations and mobility profiles. The density of the observations, as well as timing, play important roles, e.g., regular patterns during peak hours are better protected than sporadic movements. Finally, our evaluation shows that both output and input perturbation offer little additional protection, unless they introduce large amounts of noise ultimately destroying the utility of the data.

George Theodorakopoulos, Reza Shokri, Carmela Troncoso et al.

Human mobility is highly predictable. Individuals tend to only visit a few locations with high frequency, and to move among them in a certain sequence reflecting their habits and daily routine. This predictability has to be taken into account in the design of location privacy preserving mechanisms (LPPMs) in order to effectively protect users when they continuously expose their position to location-based services (LBSs). In this paper, we describe a method for creating LPPMs that are customized for a user's mobility profile taking into account privacy and quality of service requirements. By construction, our LPPMs take into account the sequential correlation across the user's exposed locations, providing the maximum possible trajectory privacy, i.e., privacy for the user's present location, as well as past and expected future locations. Moreover, our LPPMs are optimal against a strategic adversary, i.e., an attacker that implements the strongest inference attack knowing both the LPPM operation and the user's mobility profile. The optimality of the LPPMs in the context of trajectory privacy is a novel contribution, and it is achieved by formulating the LPPM design problem as a Bayesian Stackelberg game between the user and the adversary. An additional benefit of our formal approach is that the design parameters of the LPPM are chosen by the optimization algorithm.