Intrusion Prevention and Detection in Grid Computing - The ALICE Case

TL;DR

This paper proposes a new security framework for grid computing environments like CERN's ALICE experiment, combining sandboxing and machine learning-based intrusion detection to enhance security against complex attacks.

Contribution

It introduces an integrated security framework with sandboxing and machine learning for intrusion detection in grid computing, addressing current lack of comprehensive solutions.

Findings

Framework enables execution in sandboxed environments.

Machine learning detects intrusions even with zero-day attacks.

Plan to implement and test as part of ALICE Grid middleware.

Abstract

Grids allow users flexible on-demand usage of computing resources through remote communication networks. A remarkable example of a Grid in High Energy Physics (HEP) research is used in the ALICE experiment at European Organization for Nuclear Research CERN. Physicists can submit jobs used to process the huge amount of particle collision data produced by the Large Hadron Collider (LHC). Grids face complex security challenges. They are interesting targets for attackers seeking for huge computational resources. Since users can execute arbitrary code in the worker nodes on the Grid sites, special care should be put in this environment. Automatic tools to harden and monitor this scenario are required. Currently, there is no integrated solution for such requirement. This paper describes a new security framework to allow execution of job payloads in a sandboxed context. It also allows process behavior monitoring to detect intrusions, even when new attack methods or zero day vulnerabilities are exploited, by a Machine Learning approach. We plan to implement the proposed framework as a software prototype that will be tested as a component of the ALICE Grid middleware.

Figures

Tables

Table 1. Table 1: Comparison of security provided on VM and Containers.

\brVirtual Machines	Containers
\mrMore isolation layers	The kernel is directly exposed
Big surface of attack	Less mature technology
Alone, it does not solve the Grid requirements	Reduced surface of attack
	Attenuation of kernel exposition possible
	Less time to update (kernel bugs)
	Fine-grained control
\br

Full text

1 Introduction

ALICE (A Large Ion Collider Experiment) is a dedicated Pb-Pb detector designed to exploit the physics potential of nucleus-nucleus interactions at the Large Hadron Collider at CERN [1, 2]. The ALICE experiment has developed the ALICE production environment (AliEn), which implements many components of the Grid computing technologies that are needed to analyze data. Through AliEn, the computer centers that participate in ALICE can be seen and used as a single entity - any available node executes jobs and file access is transparent to the user, wherever in the world a file might be[3].

Computing Grids allow the submission of user developed jobs composed by code and data. They interface with Internet and other communication networks, also with storage systems and experiment infrastructure. They have challenging security requirements. Even when Grid system administrators perform a careful security assessment of sites, worker nodes, storage elements and central services, an attacker could still take advantage of unknown vulnerabilities (zero day). This attacker could enter and escalate her access privileges to misuse the computational resources for unauthorized or even criminal purposes. She could even manipulate the experimental data.

In this document we focus on protecting and monitoring the job execution environment. Grids require tools to enforce the environment in such a way that user processes cannot access sensitive resources. They also require automatic tools to monitor job behavior. These tools should analyze data generated in job runs like log entries, traces, system calls, to detect attacks on the system and react accordingly (for example sending alerts and stopping suspicious processes). This piece of software could be classified as a Grid Intrusion Detection System (Grid-IDS). Traditional IDS perform attack detection by fixed if-then rules based on signatures. This strategy fails when innovative intrusion methods are used. We propose the use of Machine Learning (ML) to overcome this inconvenience. Currently there is no framework that provides a solution for all above requirements. This documents provides a design for such framework.

This document is organized as follows: section 2 introduces a Grid threat model. Section 3 explains a methodology to provide security by isolation in the Grid. In section 4 there is a detailed explanation of the Grid intrusion detection strategy we plan to apply. Section 5 summarizes the project current state and the challenges faced in the design and implementation of the desired methodology. Finally, section 6 gives conclusions regarding the proposal.

2 Grid Threat model

A threat model is useful to define a set of security requirements and scope for a computing system. Threat modeling is a procedure for optimizing system security by identifying objectives and vulnerabilities, and then defining countermeasures to prevent, or mitigate the effects of threats to the system [4]. The following subsections detail a threat model for Grid computing.

2.1 The goal of the adversary

We call ’adversary’ the individual that performs unauthorized actions to increase its assigned authorization and privileges, in order to misuse the system resources or information [5, 6]. In the Grid the adversary may have one or more goals, such as:

• •

Steal sensitive data as private encryption keys, users certificates, tokens.

• •

Compromise users machines to distribute malware and steal valuable user information.

• •

Carry out a Denial of Service attack.

• •

Abuse the Grid computational resources for criminal or not allowed activities, for instance to deploy botnets or mine crypto-coins.

• •

Damage the organization reputation by using resources to attack other organizations.

2.2 Capabilities of the adversary

To achieve the above mentioned goals, an attacker could take advantage of several methods, such as:

• •

Exploit software/hardware unknown or unfixed vulnerabilities.

• •

Listen to user network to gather sensitive clear text information.

• •

Perform a ’man in the middle’ attack.

• •

Tamper other user jobs.

• •

Escalate his privileges.

• •

Access sensitive server configuration data.

2.3 Specific Grid issues to be addressed

The security requirements in Grid computing are complex and challenging. We will focus in a set of issues related to the job execution environment inside the worker nodes. Commonly, jobs running in the worker nodes have far more access rights than required, sometimes restricted only by a machine local user account. Furthermore, if all the jobs are executed with the same user, it is easy for an attacker to tamper another user job. Usually these processes have access to sensitive server data (like /etc/passwd). Currently there is not an automatic method to prevent and detect intrusions when a job is attacking the Grid. Figure 1 shows a schema of a common Grid execution environtment.

3 Intrusion prevention

We call intrusion prevention the hardening mechanism for the Grid execution environment. This mechanism guarantees that user processes cannot access privileged resources in the Grid site servers, such as server configuration files, other user jobs, the pilot job credentials and so forth. To accomplish this, we will use a security by isolation approach.

3.1 Security by isolation

Security by isolation is a technique that enforces component separation (hardware or software) in such a way that if one of them is compromised, other components remain safe [7]. Several technologies provide security by isolation, for instance Virtual Machines (VM), containers or even the Unix multiuser scheme. They provide various levels of isolation but the methods and constraints differ.

We pretend to enforce Grid site components so they run in a sandboxed environment, with Grid unprivileged users for job execution. To achieve such goal, we propose the usage of containers. If needed, they still allow like-root access inside. We could use tools like gLExec [8] as a complement. The next section gives further details about this technology. Figure 2 shows a scheme of the planned architecture of a worker node.

3.2 Containers

A container is a set of processes running on top of a common kernel (normally a Linux kernel) [9]. They are isolated from the rest of the machine and cannot affect the host or other containers. This technology uses namespaces to have a private view of the system (network interfaces, PID tree, mount points). It also uses Cgroups to have limited resource assignment. Containers can be seen as an extension of virtual memory space to a wide system scope.

Containers provide a set of features that have advantages over other virtualization technologies:

• •

They are lightweight and fast.

• •

Virtual environments.

• •

Boot in milliseconds.

• •

Just a few MB of intrinsic disk/memory usage.

• •

Bare metal performance is possible.

Figure 3 shows how containers work together, isolated and sharing the same kernel. In opposition, VM have several kernels on top of the hypervisor.

When using containers for isolation, a security impact analysis is required. Commonly, VM are used in Grid and Cloud computing to achieve isolation. Then, a comparison of provided level of isolation between containers and Virtual Machines is useful. Table 1 summarizes the main security characteristics of both alternatives. It is not clear enough which alternative provides a higher level of isolation. However, container performance characteristics and its comparable level of security makes it a suitable selection for Grid computing. Moreover, it is possible to reduce the probability that a successful attack breaks the container isolation. This will be explained in the next section.

3.3 Reducing the surface of attack

There are several alternatives to reduce the container surface of attack, decreasing the probability that a process escapes the enforced environment:

• •

Use unprivileged user and containers [10].

• •

Use Seccomp-bpf to filter available system calls [11]. This type of technology is currently being used in sanboxing methods for the Tor project, Firefox and Chrome browsers, and others.

• •

Use LSM technologies like Appamor [12].

• •

Optionally use Grsecurity Linux kernel patch [13].

• •

Optionally we could still use containers over on top of Virtual Machines.

We plan to research on these topics to find out the best solution with the least performance overhead.

4 Intrusion detection

Even if we assumed that our execution environment is perfectly safe and it is impossible to scape the enforced isolation, there are many possible attacks that can affect the Grid. It is necessary yet to analyze the job behavior to determine when an attack on the system is occurring. If an user job is misbehaving, the proposed framework should raise an alarm and perform predefined actions, for instance dropping the malicious processes. To detect innovative attacks we will use a Machine Learning algorithm. Figure 4 shows the desired implementation of the proposed system regarding Intrusion Detection.

4.1 Security metrics

There are several alternative metrics for job behavior analysis. Among them we can list the following:

• •

Job and system logs.

• •

System call sequence.

• •

Common monitoring data.

The goal is to select the best metrics that maximize relevant information about the job behavior without an evident decrease in performance. To analyze the expected huge amount of information collected, intelligent algorithms are necessary. This allows to understand in a efficient fashion the security related events happening in the Grid.

4.2 Machine Learning

For job metric analysis we propose the usage of ML techniques. Common Intrusion Detection Systems use fixed rules and search for known attack signatures. However they have problems when unknown or slightly different intrusion methods are used. ML algorithms help to analyze big amount of data by learning the expected behavior and identifying abnormal situations. we plan to gather job metrics and classify them as ’normal’ or ’malicious’. We will explore several alternative Machine Learning techniques, for instance Support Vector Machines (SVM) have been used successfully in security applications [14, 15].

5 Status

The project described in this document is currently in an early stage. At the moment we have deployed a full AliEn Grid installation in a single server. We have made several modifications to the framework to allow job execution inside unprivileged containers [10]. There are several steps still to achieve. First we should create a custom site for security testing. Then a modification of AliEn/JAliEn to fully execute jobs in containers is necessary. We will research on ML methods for IDS and finally develop a complete prototype of the proposed framework.

5.1 Challenges

There are challenges we might face in Grid security areas. For instance we should not try to increase the provided security if it means a big performance pitfall. In this paper we are considering only internal attacks, what about external attackers? We also need innovative ways to analyze the huge amount of trace/logs data generated in a efficient way. Maybe one of the most important challenges is to reduce the amount of false positives and false negatives, since the system administrators rely in the accuracy of the security monitoring framework. Finally, for further research we need to consider what would be necessary in projects where Grid private data is required, for example in the medical industry.

6 Conclussions

Computing Grids involve challenging security scenarios. Users can execute arbitrary programs inside the organization infrastructure. Consequently this execution environment has to be hardened to limit the possibility of a successful intrusion. Containers provide security by isolation among the Grid components and the underlying system. Besides, the automatic detection of attacks launched from the jobs is a priority task required to improve the overall security. To detect even innovative attack methods, a Machine Learning technique provides generalization ability to analyze job behavior information. We propose the design and development of security framework that fulfills such requirements. A software prototype will be developed and tested as a component of ALICE Grid framework AliEn.

\ack

Authors acknowledge assistance from the ALICE offline team, specially Latchezar Betev and Costin Grigoras and the CERN security department specially Stefan Lueders and Romain Wartel. Authors also knowledge the German Federal Ministry of Education and Research (Bundesministerium fur Bildung und Forschung - BMBF) for its financial support.