Spamming Botnets: Signatures and Characteristics Yinglian Xie, Fang Yu, Kannan Achan, Rina Panigrahy, Geoff Hulten, and Ivan Osipkov. SIGCOMM, 2008. Presented by: Arnold Perez Outline Introduction Goals AutoRE Challenges Design Results Botnet characteristics Contributions Weaknesses Introduction Botnets are commonly used for profit Botnets rented out to spammers Botnets can send spam emails at a large scale Can transmit thousands of emails in a short duration Difficult to detect and blacklist individual bots Goals Understand the behaviors of botnets from the perspective of large email servers that are popular targets Identify botnet characteristics and trends Track sending behavior and content patterns Develop a framework (AutoRE) that identifies botnet hosts by generating botnet spam signatures from emails AutoRE Motivated by recent success of signature based worm and virus detection systems Botnet spam emails are often sent in an aggregate fashion, resulting in content prevalence similar to worm propagation Focus primarily on URLs embedded in the email AutoRE Challenges Spammers often add random, legitimate URLs to content in order to increase the perceived legitimacy of emails AutoRE Challenges Spammers use URL obfuscation techniques to evade detection AutoRE Design AutoRE Design Input Set of unlabeled email messages Output Set of spam URL signatures Complete URL string URL regular expression List of botnet host IP addresses AutoRE Design Comprised of three modules URL preprocessor Extracts URLs and other relevant fields and groups them according to web domain Group Selects URL groups with the highest degree of burstiness in sending times RegEx selector generator Extracts signatures by processing one group at a time URL Pre-Processing Extracts URL string Source server IP address Email sending time Partitions into groups based on web domains Emails from same spam campaign always advertise the same product or service from the same domain URL Group Selection Each email my belong to more than one group Use the bursty property of botnet email traffic Select group that exhibits the strongest temporal correlation across a large set of distributed senders Signature Generation and Botnet Identification Two types of signatures Complete URL based signature Regular expression signatures Signature criteria Distributed Bursty Specific Signature Generation and Botnet Identification Distributed Total number of Autonomous Systems (AS) spanned by source IP addresses must be at least 20 Bursty The set of matching URLs must be sent within 5 days Specific Complete URLs are specific by definition For regex, entropy reduction is used to test. Probability of a random string matching signature is 1/(2^90) Automatic URL Regular Expression Generation Signature Tree Construction Constructs a keyword-based signature tree where each node corresponds to a substring, with the root of the tree set to the domain name Keywords are the most frequent substrings that are both bursty and distributed Signature Tree Construction Regular Expression Generation Detailing Returns a domain specific regular expression using the keyword-based signature Generalization Returns a more general domain-agnostic regular expression by merging very similar domain-specific expressions Regular Expression Generation Datasets and Results Based on randomly sampled Hotmail email messages November 2006 June 2007 July 2007 Total of 5,382,460 sampled emails Pre-classified as either spam or non-spam by human user (not used by filter, used for validation purposes only) AutoRE Results Identified 7,721 botnet spam campaigns 580,466 spam messages 340,050 distinct botnet host IP addresses 5,916 AS AutoRE Results AutoRE Results Majority of the campaigns belong to CU category 100% increase from July 2007 when compared to Nov 2006 Spam volume increased 50% in same time period Total number of botnet IPs does not increase proportionally, suggesting that each botnet is being used more aggressively False Positive Rate Rate = non spam matching signature / total number of non spam Ability to Detect Future Spam Experiment Apply signatures derived in Nov 2006 and June 2007 to the emails collected in July 2007 Nov 2006 signatures are not useful Indicates that spam URL patterns evolve over time June 2007 signatures are highly effective RE signatures are more robust than CU signatures over time Regular Expressions vs Keyword Conjunctions Identical spam detection rates Difference is in false positive rate Domain-specific vs DomainAgnostic Signatures Generalization effectively preserves the stable structures of polymorphic URLs while removing the volatile domain substrings Botnet Characteristics Distribution of IP addresses indicate botnet menace is a global phenomenon, with China, Korea, France, and USA having significant number of IP addresses Botnet Characteristics When viewed individually, botnet hosts do not exhibit distinct sending patterns Content in email is quite different even though the target web pages are the same 50% of botnet spam campaigns have a standard deviation of less than 1.81 hours, while 90% have standard deviation of less than 24 hours. Botnet Characteristics Similar number of recipients per email Share a constant connection rate Most likely due to rate control seen in botnet software Large number of campaigns share the same domain-agnostic regular expression signatures Same botnets participating in multiple spam campaigns Contributions AutoRE, a framework that automatically generates URL signatures for spamming botnet detection Several important findings about botnet spam Botnet hosts spread across the internet No distinctive pattern when viewed individually Botnet host sending patterns Weaknesses The AutoRE system analyzes batches of emails after they are all received Would be better if we could do this in real time to stop email once a campaign has been identified and a signature created The AutoRE system needs a lot of emails to work effectively. We can’t use it on individual inboxes, it must be put between the ISP and the incoming email Weaknesses I was hoping to take the characteristics found in the paper to use in my own project Paper shows that individually you can not identify spam from botnets. The AutoRE system works on group behavior. References "Spamming Botnets: Signatures and Characteristics". Yinglian Xie, Fang Yu, Kannan Achan, Rina Panigrahy, Geoff Hulten, and Ivan Osipkov. SIGCOMM, 2008.