Deny-Spammers: Spam Blocking with a Dynamically Updated Firewall Ruleset chris tracy deeann m.m. mikula
Motivation for Paper Deeann presented a Spam BOF at LISA 2001: –generated a lot of excitement –idea to write the paper came from this –seemed like a novel approach to fighting spam –share our “good idea” –get feedback to improve the idea
Introduction We will: –detail our methods for controlling spam at a small ISP –discuss initial unsuccessful tactics –discuss the resulting development of our unique spam blocking system
Introduction We will show how our spam blocking system: –classifies hosts as probable spammers –dynamically manages a firewall ruleset –conserves system resources –effectively blocks spam
Theft of Service Lost bandwidth CPU cycles Disk space Lost time –end-users and administrators Obviously this is lost money Apparent escalation in recent years
Spam Horror Stories Hotmail states that 80% of its almost 2 billion processed messages are spam. - Lee Gomes, The Wall Street Journal “Brightmail...now records 140,000 spam attacks a day, each potentially involving thousands of messages, if not millions.” - Jennifer Lee, The New York Times
Tools for Coping with Spam Simple Mail Filters –Mail::Audit, procmail... –patches for various MTAs qmail-uce, rblsmtpd...(lots of these) –versatile spam filters spamassassin sieve bayespam
Tools for Coping with Spam Databases –Relay Blackhole Lists (RBL) lists the IPs of known spammers, open relays, dialup/DSL address pools –Razor –Pyzor –DCC - Distributed Checksum Clearinghouse
Trouble with Spam Filtering Requires: –more CPU power –network bandwidth if accessing networked databases RBL, DCC, Razor, Pyzor... –a more complicated mail system administrative overhead –updates, configuration...
Trouble with Spam Filtering False positives (Type I error): –legitimate messages that are marked as spam –spam filters are getting better and better, but... –users are more likely to ignore everything that their spam filter catches
Our Problem Spammers were effectively DOS’ing our mail server Wanted a way to be able to selectively deny hosts –take away the ability to connect to our mail server if we detect them as a spammer –RFC 706 (next slide)
RFC On the Junk Mail Problem In a nutshell: –No mechanism for a mail host to selectively refuse messages –Lots of unwanted messages by a misbehaving host would constitute a DoS attack –Both local users and network communication could suffer
Hardware/Software Platform Software: –FreeBSD server is just a little behind... :) –qmail-1.03 patched with qmail-uce checklocal patch Hardware: –1GHz Athlon processor –640MB of RAM
What We Tried First qmail-uce checklocal patch –denys mail for non-existent mailboxes –by default, qmail accepts mail for these users this is actually an anti-spam mechanism to prevent spammers from getting valid addresses –qmail-smtpd returns a 550 when attempting to send
What We Tried First qmail-uce checklocal patch –example: RCPT TO: 550 Sorry, no mailbox here by that name. (#5.1.1)
What We Tried First qmail-uce checklocal patch –example of logging output: Oct 10 13:09:24 mail smtpd: : DENYMAIL: RCPT_TO:_Filter.NoUser:_ relay unknown [ ] FROM ADDR
What We Tried First qmail-uce checklocal patch –limitations: kept the queue size down, but didn’t prevent spammers from making parallel SMTP connections sluggish performance when there were many parallel SMTP connections to the server as a result, load average still >> 1
What We Tried First rblsmtpd (part of ucspi-tcp) –a.k.a. tcpserver – –queries any number of RBL sources (and anti-RBL sources) to catch spammers open relays, dialup/DSL pools, known spammers –temporary (451) or permanent (553)
What We Tried First rblsmtpd (part of ucspi-tcp) –toggled on when heavily spammed off if queue size < 2000 on if queue size > 2000 this was done to limit complaints –on/off method made 451 errors effectively useless –limitations: too many false positives many complaints from customers about mail delays (451) or bounces (553)
A Decision to Start Coding Should we buy more hardware? –more expensive Or write software to manage a firewall based on the checklocal logging output? –less expensive –Obviously, this is what we opted to do
Design Goals What has not worked for us in the past? Do we have enough resources to allow client-side filtering options? Do we have the time and expertise to create our own spam blocking solutions? Would it be more effective to purchase faster and better hardware than to script a custom solution? How transparent does the spam blocking need to be to the user base? Are we concerned with bandwidth consumed by spam attacks?
Requirements Method must conserve system resources. Method must reduce the amount of bandwidth consumed by spam attacks. Method must not add much additional overhead to mail processing. Method must prevent spamming sites from getting mail into the mail queue. The system must be manageable in a way that allows us to exempt certain hosts or networks. Keep our customers happy by minimizing the number of false positives. The process must be as transparent as possible to end users.
Data Flow Diagram
Oct 10 13:09:24 mail smtpd: : DENYMAIL: RCPT_TO:_Filter.NoUser:_ relay unknown [ ] FROM ADDR
Data Flow Diagram Add rule: ipfw add 1000 unreach filter-prohib tcp from to any 25 Delete rule: ipfw delete 1000
Data Structures 3 hash structures: –Host Tracking: %spammer ( hash of lists) keys of hash - host IP address values of hash - lists of timestamps timestamps - represent times that a host sent a mail to a nonexistent address –Banned Hosts: %banned (1-level hash) keys of hash - host IP address values of hash - timestamp for when a host was banned –Exception List: %noban_list (4-level hash) keys represent octets first level - first set of octets, etc... hash structure chosen for performance
Other Configuration Variables $MAX_SPAMMER_ENTRIES (default: 50) –number of timestamp entries to keep for each spammer $SPAM_TIMESPAN (default: 3600 seconds) –5-minute sampling interval –timespan to check for spam attempts $SPAM_TRIGGER (default: 10) –number of nonexistent mailbox delivery attempts required to trigger block $BAN_TIME (default: 3 days) –how long a host should stay banned for
Implementation / Pseudo-code While (true) { match maillog lines against a regexp for undeliverable messages to non-existent addresses and parse timestamp and IP address skip line if host is in the exception list trim the timestamp list for this host to $MAX_SPAMMER_ENTRIES add the timestamp to the host's list contained in the %spammer hash check how many delivery attempts to non-existent address this host has made in the sampling interval $SPAM_TIMESPAN if (nondeliverable messages count > $SPAM_TRIGGER) { add_firewall_rule(); } if (time() >= $next_refresh) { $next_refresh = time() + $REFRESH_INTERVAL; reload the exception list into %noban_list hash; prune %banned hash (un-ban hosts who have been banned for $BAN_TIME); }
In Production Firewall Reset
Quick Historical Statistics 25,284 Dec ,338 Jan ,849 Feb ,652 Mar ,175 Apr ,808 May ,298 Jun ,787 Jul ,781 Aug ,883 Sep ,935 Oct 2002 Number of hosts banned by month:
Limitations CIDR notation not supported in exception list only compatible with FreeBSD + checklocal patched qmail limited scalability checklocal exploitable by spammers to find valid addresses –easy to work around this
Future Plans Address scalability issues –add ability to use a separate firewall Integration with a 3rd-party app –SpamAssassin, Anomy Sanitizer... –use results from app to ban hosts Improve statistics generation –facilitate research –look for interesting patterns
Future Plans Develop a better interface... –for unbanning hosts –managing the exception list Interoperability with other operating systems and MTAs Develop more spam signatures...? –# of concurrent SMTP connections –# of recipients in RCPT TO list
Availability Deny-Spammers is freely available source code and documentation: – written in Perl 5 only works with FreeBSD + checklocal patched qmail
The End Thanks! Q&A...?