Section: Research Program

Attack Detection

An attack has several phases. A first major phase is the approach phase, during which the attacker enters the system, locates the target and makes himself persistent, the attack is at this point a simple intrusion. In a second phase, the attack is actually launched.

The main objective of intrusion detection is to be able to detect the attacker during the first approach phase. For that purpose, an intrusion detection system (IDS) is based on probes that continuously monitor the system. These probes generate low level alerts (warnings) for any observation of an event that could be a sign of an intrusion. These low-level alerts are very numerous and their semantic value is low. In other words, an IDS generates a huge amount of low-level alerts that bring only few information and overwhelm the security analyst. In addition, many of these alerts are actually false positives, i.e. alerts raised when there is no real intrusion.

However, these low-level alerts can themselves be considered as security events by a higher-level IDS: an alert correlation system. These higher-level IDS seek to exploit known relationships between low-level alerts to generate meta-alerts with greater semantic value, i.e. with higher-level meaning. An alert correlation system allows to reduce the number of alerts (and especially, false positives) and to return to the security analysts a higher level analysis of the situation.

There are mainly two approaches to detect intrusions. The misuse-based detection and the anomaly-based detection. A misuse-based detection is actually a signature-based detection approach: it allows to detect only the attacks whose signature is available. From our point of view, while useful in practice, misuse-detection is intrinsically limited. Indeed, it requires to continuously update the database of signatures. We follow the alternative approach, namely the anomaly approach, which consists in detecting any deviation from a reference behavior. The main difficulty is thus to compute a model of this reference behavior. Such a model is only useful if it is sufficiently accurate. Otherwise, if the model is an over-approximation, it will be a source of false negatives, i.e. real intrusions not detected. If the model is a under-approximation, it will be a source of false positives, i.e. normal behaviors seen as intrusions.

In this context, our contributions in intrusion detection systems follow two separate axes: anomaly-based IDS and alert correlation systems. Our contribution in anomaly-based intrusion detection relies on:

  • Illegal Information Flow Detection: we have proposed to detect information flows in the monitored system (either a node or a set of trusted nodes) that are allowed by the access control mechanism, but are illegal from the security policy point of view. This approach is particularly appealing to detect intrusions in a standalone node.

  • Anomaly-Based Detection in Distributed Applications: our goal is to specify the normal behavior based on either a formal specification of the distributed application, or previous executions. This approach is particularly appealing to detect intrusions in industrial control systems since these systems exhibit well-defined behaviors at different levels: network level (network communication patterns, protocol specifications, etc.), control level (continuous and discrete process control laws), or even the state of the local resources (memory or CPU).

  • Online data analytics: our goal is to estimate on the fly different statistics or metrics on distributed input streams to detect abnormal behavior with respect to a well-defined criterion such as the distance between different streams, their correlation or their entropy.