Projects

C5: Security in Invasive Computing Systems

Principal Investigators:

Prof. F. Freiling, Prof. W. Schröder-Preikschat, Prof. G. Snelting,
in cooperation with Mercator Fellow Prof. I. Verbauwhede

Scientific Researchers:

S. Bischof, F. Schirrmacher, F. Turan

Abstract

Project C5 explores security aspects of invasive computing and resource-aware programming. We aim to ensure confidentiality, integrity and availability of the invasive computing system in the presence of untrustworthy programs that compete for resources and can contain malicious functionality, thereby closing a gap in the system architecture. Project C5 has been established within the second funding phase.

In the second funding phase, the groundwork for security concepts had to be laid: We began to devise the attacker model and define specific security properties for invasive applications. Specific to invasive computing, the attacker model comprises an absent or untrusted system layer accessing shared resources. We then devised several new mechanisms that achieved different levels of isolation at different architectural layers against attacker models with different assumption coverage: at the hardware layer (i.e. isolating applications running on the same core against system level attackers), systems software layer (isolating memory abstractions against application level attackers), and the application layer (integrating security requirements as constraints into the invade phase of an invasive program).

In the third funding phase, we aim to pursue two main goals: (1) We wish to provide mechanisms that can reliably and provably enforce security properties that have been requested by applications at run time even if attacker assumptions are violated, thereby increasing the assumption coverage. (2) We plan to finally extend intra-tile to inter-tile security. The main idea that contributes to both goals is to combine static information flow control with remote attestation together with a hardware-based remote (and destructive) reset mechanism to check and (re-)enforce integrity of computations on remote tiles. Ideas specifically in the direction of goal 1 are heuristic run-time-monitoring approaches and static quantitative information flow analyses to detect information leaks within the invasive system. The second goal specifically is targeted by a novel application of memory encryption on inter-tile virtual shared memory.

Synopsis

The overall challenge is always that security is only as strong as the weakest link: Even though we have gained many insights into security considerations on invasive computing systems already, many challenges for security at different abstraction layers remain. Security holes can exist at the application layer, at different software layers, the architecture layer down to the hardware layer. Security attacks can also take on many forms and shapes. All this needs to be taken into account.

Building upon the results of the second phase, we formulate two main goals for our investigations in the third phase:

• Goal 1: Provide robust run-time monitoring and enforcement mechanisms for security properties to increase assumption coverage.
• Goal 2: Extend intra-tile to inter-tile security for integrity and confidentiality.

The first goal aims at integrating the diverse solutions for the different layers that were developed in previous work. The choice of layer on which defensive techniques are placed critically depends on the attacker assumption. The choice of attacker assumption, however, is usually performed by the system operator who can misjudge the threat scenario. The idea which we pursue is to develop solutions that enable the system to detect the violation of attacker assumptions and enable more restrictive security techniques automatically. This is achieved through robust combinations of techniques already developed in the second phase or being developed in this or other projects of the SFB in the third phase. Overall, the probability of choosing the right attacker assumption (and therefore the assumption coverage) is increased.

The second goal is directed towards bootstrapping security from single tiles to multiple tiles. Regarding the hardware layer, the focus in the second phase has been to develop stand-alone solutions (individual cores) that can act as trust anchors in an invasive system. If strong attackers are present, it is not sufficient to maintain islands of security; ideally it should be possible to reset applications that do not obey their prescribed control flow from within the trust anchors. Ideally, the entire system can be brought into a well-defined state again.

Approach

We wish to explore possible combinations of multiple existing trusted computing solutions (those developed in the second phase and others) to form more powerful security mechanisms in heterogeneous multicore architectures. This approach is relevant to both goals formulated above. We now describe the main ideas and concepts behind our solutions.

One main thrust will be to compose solutions for control flow attestation with static information flow analysis to achieve a novel and strong form of information flow protection against strong adversaries. This idea contributes mainly to Goal 2 but also to Goal 1, since violations of integrity can then be remotely detected. Hardware-based mechanisms to remotely force non-behaving applications to recover from adversarial actions through a novel concept of destructive preemption are investigated as well. Techniques that ensure inter-tile integrity and thus support Goal 2 at the operating system level are explored, where virtual shared memory (VSM) is used to achieve inter-tile encryption, integrity and authorisation against intermediate strength adversaries. At the same level, we wish to study run-time monitoring techniques that can heuristically detect confidentiality violations.

For example, we wish to introduce static information flow control (IFC) on the programming language level for invasive applications to guarantee confidentiality and integrity in close collaboration with C3. Here, confidentiality means that secret inputs do not influence public outputs, while integrity means that untrusted inputs cannot influence trusted behaviour. Our work will leverage the JOANA system, which uses a new, groundbreaking algorithm for probabilistic non-interference and provides IFC for full multithreaded Java (an X10 front end is in preparation).

While we currently focus solely on static attestation, i.e. the authenticity and integrity of an application is only checked while being loaded, we plan to leverage our solutions to support control flow attestation (CFA). Without CFA, the integrity property of a program can be violated, for example because implementation bugs are exploited and the contents of application memory are changed after the application has been loaded. We will address this problem by extending our static attestation solutions such as Soteria and Atlas with CFA but instead of extending measurements on a basic block basis, we only want to consider basic blocks which belong to critical sections within an applications control flow graph. To this end, we will utilise the static IFC algorithms provided by JOANA such that only critical points (see above) need to be checked and measured during run time while the other parts of the application are considered secure.

Finally, we aim to equip each tile with a hardware extension capable of measuring available software and data, supporting attestation to neighbouring tiles. These tile-to-tile attestations will be used to measure whether these tiles can be trusted to process sensitive workloads, or if the code running there should be revised first. For example, a long-running application may require patches to protect against newly detected attack vectors. Consequently, these mutual attestations will ensure that every invaded tile has the most trusted version of the software or data, i.e. the most recent version addressing all known vulnerabilities. Moreover, possible actions in such an attestation policy include updates, even when doing so requires preemption of the running system to launch into the inherently trusted execution environment which can forcibly apply the update.

A comprehensive summary of the major achievements of the second funding phase can be found by accessing Project C5 second phase website.

Publications