Today large software systems (i.e., giants) thrive in commodity markets, but are untrustworthy due to their numerous and inevitable software bugs that can be exploited by the adversary. Thus, the best hope of security is that some small, simple, and trustworthy software components (i.e., wimps) can be protected from attacks launched by adversary-controlled giants. However, wimps in isolation typically give up a variety of basic services (e.g., file system, networking, device I/O), trading usefulness and viability with security. Among these basic services, isolated I/O channels remained an unmet challenge over the past three decades. Isolated I/O is a critical security primitive for a myriad of applications (e.g., secure user interface, remote device control). With this primitive, isolated wimps can transfer I/O data to commodity peripheral devices and the data secrecy and authenticity are protected from the co-existing giants. This thesis addresses this challenge by proposing a new security architecture to provide services to isolated wimps. Instead of restructuring the giants or bloating the Trusted Computing Base that underpins wimp-giant isolation (dubbed underlying TCB), this thesis presents a unique on-demand I/O isolation model and a trusted add-on component called wimpy kernel to instantiate this model. In our model, the wimpy kernel dynamically takes control of the devices managed by a commodity OS, connects them to the isolated wimps, and relinquishes controls to the OS when done. This model creates ample opportunities for the wimpy kernel to outsource I/O subsystem functions to the untrusted OS and verify their results. The wimpy kernel further exports device drivers and I/O subsystem code to wimps and mediates the operations of the exported code. These two methodologies help to significantly reduce the size and complexity of the wimpy kernel for high security assurance. Using the popular and complex USB subsystem as a case study, this thesis illustrates the dramatic reduction of the wimpy kernel; i.e., over 99% of the Linux USB code base is removed. In addition, the wimpy kernel also composes with the underlying TCB, by retaining its code size, complexity and security properties.

Cette thèse souhaite fournir des outils pour qu’un utilisateur puisse contribuer activement à la sécurité de son usage d’un système informatique. Les activités de sensibilités différentes d’un utilisateur nécessitent tout d’abord d’être cloisonnées dans des domaines dédiés, par un contrôle d’accès s’ajustant aux besoins de l’utilisateur. Afin de conserver ce cloisonnement, celui-ci doit être en mesure d’identifier de manière fiable les domaines avec lesquels il interagit, à partir de l’interface de sa machine. Dans une première partie, nous proposons un nouveau mécanisme de cloisonnement qui peut s’adapter de manière transparente aux changements d’activité de l’utilisateur, sans altérer le fonctionnement des contrôles d’accès existants, ni dégrader la sécurité du système. Nous en décrivons une première implémentation, nommée StemJail, basée sur les espaces de noms de Linux. Nous améliorons ce cloisonnement en proposant un nouveau module de sécurité Linux, baptisé Landlock, utilisable sans nécessiter de privilèges. Dans un second temps, nous identifions et modélisons les propriétés de sécurité d’une interface homme-machine (IHM) nécessaires à la compréhension fiable et sûre du système par l’utilisateur. En particulier, il s’agit d’établir un lien entre les entités avec lesquelles l’utilisateur pense communiquer, et celles avec lesquelles il communique vraiment. Cette modélisation permet d’évaluer l’impact de la compromission de certains composants d’IHM et d’aider à l’évaluation d’une architecture donnée. / This thesis aims to provide end users with tools enhancing the security of the system they use. First, user activities of different sensitivities require to be confined in dedicated domains by an access control fitting the user’s needs. Next, in order to maintain this confinement, users must be able to reliably identify the domains they interact with, from their machine’s interface. In the first part, we present a new confinement mechanism that seamlessly adapts to user activity changes, without altering the behavior of existing access controls nor degrading the security of the system. We also describe a first implementation named StemJail, based on Linux namespaces. We improve this confinement tool by creating a new Linux security module named Landlock which can be used without requiring privileges. In a second step, we identify and model the security properties a human-computer interface (HCI) requires for the reliable and secure understanding of the system by the user. Precisely, the goal is to establish a link between the entities with which the users think they communicate, and those with which they actually communicate. This model enables to evaluate the impact of HCI components jeopardization and helps assessing a given architecture.