Protecting Privacy in Multi-agent Optimization

Many real-life optimization problems involve multiple entities, or agents (individuals, companies...), with their own private constraints and preferences, communicating with each other in order to find a solution that maximizes the overall public satisfaction. In Artificial Intelligence, the field of Distributed Constraint Optimization (DCOP) has been addressing such multi-agent optimization problems, through distributed message-passing algorithms such as the DPOP algorithm. However, the research in DCOP has been neglecting the privacy of the information exchanged by the agents during the computing of the solution, which is critical to many real-life problems. While agents are willing to cooperate with each other to produce an optimal solution, they are most often reluctant to reveal their private constraints and preferences to other, which hinders this cooperation. The goal of this project was to implement, test, and evaluate a secured version of DPOP, P-DPOP. P-DPOP provides strong agent privacy and topology privacy by randomization, constraints privacy and limited decision privacy by obfuscation. The algorithm was implemented in Java, as part of the open-source FRODO platform for DCOP.

Scalable Multi-agent Coordination and Resource Sharing

Panayiotis Danassis

A plethora of real world problems consist of a number of agents that interact, learn, cooperate, coordinate, and compete with others in ever more complex environments. Examples include autonomous vehicles, robotic agents, intelligent infrastructure, IoT devices, and so on. As more and more autonomous agents are deployed in the real-world, it will bring forth the need for novel algorithms, theory, and tools to enable coordination on a massive scale. In this thesis, we develop such tools to tackle two central challenges in multi-agent coordination research: solving allocation problems, and resource sharing, focusing on solutions that are scalable, practical, and applicable to real-world problems.In the first part of the thesis we tackle the problem of allocating resources to agents, i.e., solving a weighted matching problem. Real-world matching problems may occur in massively large systems, they are distributed and information-restrictive, and individuals have to reveal their preferences over the possible matches in order to get a high quality match, which brings forth significant privacy risks. As such, there are three main challenges: complexity, communication, and privacy.Our proposed approach, ALMA, is a practical heuristic designed for real-world, large-scale (

10^6

agents) applications. It is based on a simple altruistic behavioral convention: agents have a higher probability to back-off from contesting a resource if they have good alternatives, potentially freeing the resource for some agent that does not. ALMA tackles all of the aforementioned challenges: it is decentralized, runs on-device, requires no inter-agent communication, converges in constant time -- under reasonable assumptions --, and provides strong, worst-case, privacy guarantees. Moreover, by incorporating learning we can mitigate the loss in social welfare and increase fairness. Finally, rational agents can use such simple conventions, along with an arbitrary signal from the environment, to learn a correlated equilibrium for accessing a set resources, under high congestion.In the second part of the thesis we focus on a critical open problem: the question of cooperation in socio-ecological and socio-economical systems, and sustainability in the use of common-pool resources. In recent years, learning agents, especially deep reinforcement learning agents, have become ubiquitous in such systems. Yet, scaling to environments with a large number of agents and low observability continues to be a challenge. In our work, we focus on common-pool resources. Individuals face strong incentives to appropriate, which results in overuse and even the depletion of the resources. Our goal is to apply simple interventions to steer the population to desirable states.We propose a simple, yet powerful, and robust technique: allow agents to observe an arbitrary common signal from the environment. The agents learn to couple their policies, and avoid depletion in a wider range of settings, while achieving higher social welfare and convergence speed.Finally, we propose a practical approach to computing market prices and allocations via a deep reinforcement learning policymaker agent. Compared to the idealized market equilibrium outcome -- which can not always be efficiently computed -- our policymaker is much more flexible, allowing us to tune the prices with regard to diverse objectives such as sustainability and resource wastefulness, fairness, buyers' and sellers' welfare, etc.

EPFL2022

Protecting privacy through metadata analysis

Sandra Deepthy Siby

Although encryption hides the content of communications from third parties, metadata, i.e., the information attached to the content (such as the size or timing of communication) can be a rich source of details and context. In this dissertation, we demonstrate the power of metadata analysis. We illustrate ways in which we can use metadata analysis to protect privacy, by addressing two problems in the areas of network and web privacy.In the first problem, we study recently standardized protocols such as encrypted DNS and QUIC. We show how metadata analysis can be used by adversaries to perform website-fingerprinting attacks against these protocols and to infer websites visited by a user, thereby compromising their privacy. We use the insights from our analysis to identify the requirements for developing effective countermeasures and to improve the resistance of these protocols to website fingerprinting. We find that hiding metadata is challenging.In the second problem, we address the issue of online advertising and tracking services (ATS) that are constantly evolving to evade privacy protections established by browser vendors. We demonstrate how the very fact of metadata being hard to hide can be useful to defenders. Defenders can use metadata analysis to build ATS-detection systems that are more robust against adversarial evasion by capturing behavioral metadata of ATS. We use our findings to detect and counter the emergence of tracking via first-party cookies.

EPFL2022

Distributed Constraint Optimization

Thomas Léauté

Distributed Constraint Satisfaction (DisCSP) and Distributed Constraint Optimization (DCOP) are formal frameworks that can be used to model a variety of problems in which multiple decision-makers cooperate towards a common goal: from computing an equilibrium of a game, to vehicle routing problems, to combinatorial auctions. In this thesis, we independently address two important issues in such multi-agent problems: 1) how to provide strong guarantees on the protection of the privacy of the participants, and 2) how to anticipate future, uncontrollable events. On the privacy front, our contributions depart from previous work in two ways. First, we consider not only constraint privacy (the agents' private costs) and decision privacy (keeping the complete solution secret), but also two other types of privacy that have been largely overlooked in the literature: agent privacy, which has to do with protecting the identities of the participants, and topology privacy, which covers information about the agents' co-dependencies. Second, while previous work focused mainly on quantitatively measuring and reducing privacy loss, our algorithms provide stronger, qualitative guarantees on what information will remain secret. Our experiments show that it is possible to provide such privacy guarantees, while still scaling to much larger problems than the previous state of the art. When it comes to reasoning under uncertainty, we propose an extension to the DCOP framework, called DCOP under Stochastic Uncertainty (StochDCOP), which includes uncontrollable, random variables with known probability distributions that model uncertain, future events. The problem becomes one of making "optimal" offline decisions, before the true values of the random variables can be observed. We consider three possible concepts of optimality: minimizing the expected cost, minimizing the worst-case cost, or maximizing the probability of a-posteriori optimality. We propose a new family of StochDCOP algorithms, exploring the tradeoffs between solution quality, computational and message complexity, and privacy. In particular, we show how discovering and reasoning about co-dependencies on common random variables can yield higher-quality solutions.

EPFL2011

Distributed Constraint Optimization

Thomas Léauté

EPFL2011

Scalable Multi-agent Coordination and Resource Sharing

Panayiotis Danassis

10^6

EPFL2022