Detecting Anomalies in Inter- hosts Communication Graph Jan, 14, - PowerPoint PPT Presentation

Detecting Anomalies in Inter- hosts Communication Graph Jan, 14, 2009 Keisuke ISHIBASHI*, Tsuyoshi KONDOH*, Shigeaki HARADA § , Tatsuya MORI § , Ryoichi KAWAHARA § , Shoichiro ASANO ¶ *NTT Information Platform Labs. § NTT Service Integration Labs. ¶ National Information Institute Flocon2009 1

Outline • Anomalous traffic detection • Inter-host communication graph • Anomalies in communication graph • Detecting method for graph anomaly – Similarities between graphs • Experimental results – Synthesized traffic – Actual traffic Flocon09 2

Anomalous traffic detection • DDoS attacks, Network failure etc: can be detected as sudden change in traffic volume • Worm scans or botnet C&C traffic: cannot be found as volume change – Whose traffic volume is very small, and buried in normal traffic • May be found as sudden change in traffic pattern, not volume • Traffic pattern – Entropy: can reveal traffic characteristic per hosts. – Communication pattern between hosts: can reveal anomalous traffic which appears as inter-hosts communication pattern Flocon2009 3

Communication pattern between hosts • Can be represented as graph • Communication graphs for anomalous traffic – Some of them are difficult to detect with conventional methods • Conventional methods: monitoring entropies in number of flows, etc Botnet Victims Botnet Worm C&C C&C infected Victims server Victims server hosts Worm scan Botnet P2P Botnet More difficult to detect Flocon2009 4

5 Time series of communication graph Flocon2009

Challenge • How to detect anomaly (change) in time series of graph? • Visualization or animation of commutation graph[Yurcik06] – Useful especially for digging anomalous event by hand – However, eyeballing by human operator is needed to detect anomalous event • Automated detection: need to define similarity between graphs S(G t ,G t+1 ), where G t and G t+1 are graphs of time t and t+1 – Can judge as an anomaly if S(G t ,G t+1 ) suddenly decreases t=3 t=2 S(G 2, G 3 ) t=1 t=0 S(G 1, G 2 ) S(G o, G 1 ) • [Yurcik06] William Yurcik, “VisFlowConnect-IP: A Link-Based Visualization of NetFlows for Security Monitoring,” 18 th Annual FIRST Conference, June 2006. Flocon2009 6

Similarities between graphs • Graph Kernel – Define “inner product” like function f(•, •), a.k.a kernel, on the space of non-linear spaces [Kashima03] • Edit distance – Number of operations to change graph G to G’ [Bunke06] – operations: add/remove edges/nodes • Can be used to detect anomalies in graph time-series • Difficult to identify the source of anomaly • [Kashima03] H. Kashima, et.al , “Marginalized kernels between labeled graphs,” In Proc. ICML 2003, pp.321-328. • [Bunke06] H. Bunke et.al, “Computer Network Monitoring and Abnormal Event Detection Using Graph Matching and Multidimensional Scaling, ” LNCS Vol. 4065 2006. Flocon2009 7

Linear feature space projection • Linear feature space projection[Ide04] – Mapping a graph to a vector in the linear space that represents the feature of the graph • As feature vectors, adopt a principal eigenvector of adjacency matrix for the graph – ≈ Page Rank vector – Dimension of linear space: Number of nodes in graphs Host3 1 2 3 Host2 1 - 1 1 Host2 Host1 2 1 - 0 Host3 3 1 0 - Principal Host1 eigen Communication graph Feature vector Adjacency matrix vector • [Ide04] Tsuyoshi Ide and Hisashi Kashima: Eigenspace-based Anomaly Detection in Computer Systems, In Proc. 10th ACM SIGKDD Conference (KDD2004), Seattle, WA, USA, 2004. Flocon2009 8

Anomaly detection using feature vector • Periodically generate communication graph from observed traffic data, and calculate feature vectors of the graphs • Calculate similarity between the graph and the previous one Cosine similarity • Judge as anomaly if the similarity suddenly decreases High similarities Vector elements for Host3 Vector at time t Vector at time t+1 Low similarities-> detected as anomaly Host2 Vector at time t+2 Host1 Flocon2009 9

Compressing adjacency matrix • In large communication graph, calculating principal eigen vector of adjacency matrix may be difficult. • Compress adjacency matrix by combining hash matrix and bloom filter Source Address Destination Address Hash(DstIP) 192.168.0.1 → 10.0.0.1 Hashing 1 2 3 M Hashing Source- Destination Pair 1 1 1 1 1 Hash(SrcIP) 2 1 1 2 1 H(192.168.0.1.10.0.0.1) 3 1 1 0 1 Chech whether the pair BloomFIiter M 0 1 1 0 is new or not If new, then increment the corresponding cell Flocon2009 10 10

Experimental results • Observed data: packet capture data of 24-hour long at 1Gbps link • Use packets with ports 135/445(scans)/6667(IRC) – Current python implementation cannot handle whole traffic – Focus on botnet related traffic • Generate graphs every minutes • Hash matrix size ： 1280 × 1280 Flocon2009 11

Time series of simulates of feature vectors • Several sudden decreases in similarities • Try to find the source of anomaly for the first one eigv ec 1.2 1 0.8 Elapsed time 0.6 0.4 eigvec 0.2 0 0:00 3:00 6:00 9:00 12:00 15:00 18:00 21:00 0:00 Similarity Flocon2009 12

Comparison of graphs before/after the anomaly • By comparing graphs and/or vectors before/after the anomaly, we can identify the source of anomaly • Comparing vectors is fit for automated identification • In this case: sudden large virus scan 8000 1 7000 0.8 6000 5000 0.6 degvec-before 4000 eigvec-before 0.4 3000 2000 0.2 1000 0 0 0 200 400 600 800 1000 1200 1400 Flocon2009 13

Evaluation with synthesized anomaly cluster • Which type of anomaly and how large anomaly can be detected by the proposed method? • Evaluation using synthesized anomaly can answer the above question • Firstly, mesh cluster of various size is inserted to actual communication graph and calculate the similarity between the original graph Flocon2009 14

Evaluation with synthesized anomaly cluster • With mesh size > 70, similarity decreases and the anomaly can be found 1.2 1 0.8 Similarity 0.6 degvec 0.4 eigvec 0.2 0 0 20 40 60 80 100 120 Num of mesh nodes Flocon2009 15

Conclusion • Summary – Propose a method to detect anomalies in communication graphs • Projection of graph into linear feature spaces, and compare the simulates between feature vectors – Evaluate using actual traffic data • Found a sudden large worm scan • Future works – Apply to other traffic data to find out which type of anomaly the proposed method can detect – Faster implementation Flocon2009 16

Acknowledgement • This study was supported in part by the Ministry of Internal Affairs and Communications of Japan. Flocon2009 17