Cognitive Security的異常檢測技術

最近，Cisco重返網絡安全的一個標志性收購就是買下了位于捷克的Cognitive Security公司。這家由捷克一所大學老師創立的startup公司有啥看家的本領呢？呵呵，原來就是DFI，或者說是基于流量的異常檢測技術。

Cognitvie的目標很明確，就是檢測APT，還有0-day***，以及其他多態惡意代碼。

Cognitive用到了以下基于異常的檢測算法，不是什么新的算法，但是他們做到了實用化。

Cognitive Analyst's products and services utilize a multi-stage detection algorithm to generate a Cognitive Trust Score (CTS), which is effectively a measure of ''Trustfulness' to the data which is being analyzed. Currently eight stages are used to increase the detection and accuracy of threats, and collectively generate an accurate CTS for an analyst to action and subsequently mitigate against an attack. A selection of these algorithms are summarized as follows:

• MINDS algorithm [Ertoz et al, 2004] 【一種基于源/目標分析的***檢測算法】The Minnesota Intrusion Detection System (MINDS) processes data from a number of flows: 1. Data from a single source IP to multiple destinations, 2. flows from multiple sources to a single destination, or 3. a series of flows between a single source to a single destination.
• Xu et al. algorithm [Xu, Zhang et al, 2005] 【一種流量源分類算法】This algorithm serves to classify traffic sources. A normalized entropy is established (i.e. establishing meaningful analysis to the apparent randomness of a data set), determined by applying static classification rules to the established normalized states.
• Volume prediction algorithm [Lakhina et al, 2004] 【流量預測算法】uses the Principal Components Analysis (PCA) methodology, which is a mathematical procedure used to formulate predictive models. In order to build a model of traffic volumes from individual sources, values are determined based on the number of flows, bytes, and packets generated from each source. The PCA method then identifies the complex relationships between the traffic originating from distinct sources.
• Entropy prediction algorithm [Lakhina et al, 2005]【熵預測算法】 This algorithm is similar to the PCA-based traffic modeling discussed above, but uses different features than just volume prediction. Entropy prediction aggregates traffic from source IPs, but instead of processing traffic volumes, it predicts the entropy of source and destination ports, and destination IPs.
• TAPS algorithm [Sridharan et al, 2006]【一種流量逐層分析算法】 targets a specific class of attacks by classifying a subset of suspicious traffic sources and characterizing them by three features: 1. the number of destination IP addresses, 2. the number of ports in the set of flows from the source, and 3. the entropy of the flow size. The anomaly of the source is based on the ratios between these values.

其實，對于這類技術，我已經多次提到過了。我們也在這方面做出了很多努力和工作，并且也已經用到了我們的產品之中。

【參考】

基于異常的檢測技術