Transcript ppt - MMLab
Googling the Internet Unconstrained Endpoint Profiling Ionut Trestian, Supranamaya Ranjan, Alekandar Kuzmanovic, Antonio Nucci Reviewed by Lee Young Soo Introduction For understanding what people are doing on the Internet Analyze operational network trace. Obtaining ‘raw’ packet trace from operational networks can be very hard. Accurately classifying in an online fashion at high speeds is an inherently hard problem. Unconstrained Endpoint Profiling Introduction of a novel methodology. No operational traces are available Packet-level traces are available Sampled flow-level traces are available Internet access trend analysis for four world regions. Methodology Rule Generation Querying Google using a sample ‘seed set’ of random IP address from the networks in four world regions. Constrain top N keywords that could be meaningfully used for endpoint classification. Methodology Methodology Web Classifier Rapid URL search Hit text search Example URL : www.robtex.com/dns/32.net.ru.html Methodology IP tagging URL based tagging General hit text based tagging Hit text based tagging for Forums Post-date & username is in the vicinity of the IP address => forum user Presence of following keywords :http:\, ftp:\, ppstream:\, mms:\ => http share, ftp share, streaming node Methodology Examples 200.101.18.182-inforum.insite.com URL based tagging 61.172.249.13-ttzai.com Hit text based tagging for Forum Information come from Web logs Proxy logs Forums Malicious list Server list P2P communication Evaluation When No Traces are Available. When Packet-Level Trace are Available. When Sampled Trace are Available. When No Traces are Available Applying the unconstrained endpoint approach on a subset of the IP range belonging to four ISPs shown in above table. When No Traces are Available When No Traces are Available Correlation with operational traces. Correlation with other sources. Unconstrained endpoint profiling approach can be effectively used to estimate application popularity trends. When Packet-Level Trace are Available BLINC UEP Off-line tool Superior classification result Cannot classify particularly at application level Variable quality result for different traces Efficiently operate online When Packet-Level Trace are Available Collect most popular 5% of IP address and tag them by applying the methodology. Use this information to classify the traffic flow. When Packet-Level Trace are Available When Sampled Trace are Available Due to sampling, insufficient amount of data remains in the trace, and hence the graphlets approach simply does not work. Popular endpoint are still present in the trace, despite sampling. When Sampled Trace are Available Endpoint approach remains largely unaffected by sampling. Endpoint Profiling Endpoint Clustering Employ clustering in networking has been done before : Autoclass algorithm. A set of tagged IP addresses from region’s network Input to the endpoint clustering algorithm. Endpoint Profiling Browsing, browsing and chat or mail seems to be most common behavior. Endpoint Profiling Traffic Locality Conclusion UEP Accurately predict application and protocol usage trends when no network traces are available. Dramatically out perform when packet traces are available. Retain high classification capabilities when flow-level traces are available. Profile endpoints residing at four different world regions. Network applications and protocols used in these region. Characteristics of endpoint classes that share similar access patterns. Clients’ locality properties.