Backtracking Algorithmic Complexity Attacks Against a NIDS Randy Smith, Cristian Estan, Somesh Jha University of Wisconsin–Madison
Algorithmic Complexity Attacks Vulnerable algorithm: algorithm whose worst case differs from typical case. The larger the difference, the more vulnerable the algorithm. Examples: Algorithm Average Worst Quicksort O(n log n) O(n2) Hash lookup constant O(n)
Algorithmic Complexity Attacks Algorithmic Complexity Attack – an attacker induces worst-case behavior in a vulnerable algorithm. Common observable effect is denial of service. Crosby and Wallach: induced worst-case behavior in hash function implementations. “Algorithms are now part of the attack surface” (Crosby and Wallach, 2003)
Are NIDS vulnerable? NIDS and IPS are ubiquitous, but… Do they contain vulnerable algorithms? Can they be exploited? YES! Only need 1 packet every 3 seconds.
Evading a NIDS Attacker’s Goal: Evade NIDS Two attack vectors in an evasion attempt: 1st—alg. complexity attack targeting the NIDS 2nd—true attack targeting the network Effect of an algorithmic complexity attack: (NIDS) Packets enter network unexamined (fail-closed IPS) Packets are dropped
Main results In Snort, vulnerability in rule-matching worst-case vs. typical case: 6 orders of magnitude. “Backtracking Attack” Easily exploitable through packet payloads Improved rule-matching algorithm limits running time differences to within 1 order of magnitude.
Outline Snort rule matching Inducing backtracking attacks Countermeasures Measurement results Conclusion
Snort Rule Matching alert tcp $EXT_NET any -> $HOME_NET 99 content:”fmt=”; //P1 content:”player=”; //P3 content:”overflow”,relative; //P5 alert tcp $EXT_NET any -> $HOME_NET 99 (msg:”AudioPlayer jukebox exploit”; pcre:”/^(mp3|ogg)/”,relative; //P2 pcre:”/.exe|.com/”,relative; //P4 sid:5678)
Snort Rule Matching Rule matches! alert tcp $EXT_NET any -> $HOME_NET 99 (msg:”AudioPlayer jukebox exploit”; content:”fmt=”; //P1 pcre:”/^(mp3|ogg)/”,relative; //P2 content:”player=”; //P3 pcre:”/.exe|.com/”,relative; //P4 content:”overflow”,relative; //P5 sid:5678) Rule matches! fmt=acc player=default fmt=mp3 rate=14kbps player=cmd.exe?overflow#@!%
Matching the packet P1 alert tcp $EXT_NET any -> $HOME_NET 99 (msg:”AudioPlayer jukebox exploit”; content:”fmt=”; //P1 P3 pcre:”/^(mp3|ogg)/”,relative; //P2 content:”player=”; //P3 pcre:”/.exe|.com/”,relative; //P4 P4 content:”overflow”,relative; //P5 sid:5678) P5 Rule matches! fmt=acc player=default fmt=mp3 rate=14kbps player=cmd.exe?overflow#@!%
Inducing Backtracking attacks P1,P2,P3,P4 match in 3 positions each P5 never matches alert tcp $EXT_NET any -> $HOME_NET 99 (msg:”ReelAudio jukebox exploit”; content:”fmt=”; //P1 pcre:”/^(mp3|ogg)/”,relative; //P2 content:”player=”; //P3 pcre:”/.exe|.com/”,relative; //P4 content:”overflow”,relative; //P5 sid:5678) Leads to excessive packet traversals! fmt=mp3fmt=mp3fmt=mp3player=player=player=.exe.exe.exe fmt=acc player=default fmt=mp3 rate=14kbps player=cmd.exe?overflow#@!%
Matching the malicious packet alert tcp $EXT_NET any -> $HOME_NET 99 (msg:”AudioPlayer jukebox exploit”; content:”fmt=”; //P1 pcre:”/^(mp3|ogg)/”,relative; //P2 content:”player=”; //P3 pcre:”/.exe|.com/”,relative; //P4 content:”overflow”,relative; //P5 sid:5678) P2 P3 P4 P4 P5 P4 P5 P5 P5 P5 P5 P5 fmt=mp3fmt=mp3fmt=mp3player=player=player=.exe.exe.exe
Are real rules vulnerable? Rule number Processing (s/GB) Slowdown Same proto All traffic 3682 (SMTP) 30,933,874 232,936X 1,501,644X 2611 (Oracle) 6,220,768 56,296X 301,979X 1382 (IRC) 1,956,858 134,031X 94,993X 2403 (NetBIOS) 357,777 490X 17,368X 1755 (IMAP) 89,181 444X 4,329X
Safer backtracking Memoization: maintain a table of subproblem “answers”; never evaluate a predicate twice at the same starting payload offset alert tcp $EXT_NET any -> $HOME_NET 99 (msg:”AudioPlayer jukebox exploit”; content:”fmt=”; //P1 pcre:”/^(mp3|ogg)/”,relative; //P2 content:”player=”; //P3 pcre:”/.exe|.com/”,relative; //P4 content:”overflow”,relative; //P5 sid:5678) Identify constrained predicate sequences Monotone memoization: don’t re-evaluate monotone predicates that have been evaluated at lower offsets
Reductions in processing cost 4 11 18 P5 P4 P2 P3 P5 P4 P2 P3 P2 7 14 21 P3 28 35 42 P4 P4 P4 46 50 54 P5 P5 P5 P5 P5 P5 P5 P5 P5 fmt=mp3fmt=mp3fmt=mp3player=player=player=.exe.exe.exe
Outline Snort rule matching Inducing backtracking attacks Protecting against backtracking attacks Measurement results Conclusion
Slowdown factor w.r.t. same protocol Measurement results Rule number Slowdown factor w.r.t. same protocol Before w/ Memo+ 3682 (SMTP) 232,936X 0.95X 2611 (Oracle) 56,296X 1.57X 1382 (IRC) 134,031X 6.00X 2403 (NetBIOS) 490X 0.17X 1755 (IMAP) 444X 0.46X
Live experiment topology Background Traffic AC Attack True Attack
Live experiment Background Traffic @ 10Mbps AC Attack Targets Snort SMTP rule 3682 Directed at sendmail server True Attack: NIMDA 300 exploit attempts, sent 1 byte per second. New exploit started every second.
Live experiment results Attack Description Exploits Detected Required Rate (kbps) Control (No attack) 300/300 -- 2 packets every 60 s. 220/300 0.4 1 packet every 5 s. 4/300 2.4 1 packet every 3 s. 0/300 4.0 20 packets initially 0.8
Conclusions NIDS operation is complex. Many opportunities for vulnerable algorithms. In Snort, rule-matching is vulnerable and can be exploited by an attacker. Memoization, along with other semantics-preserving operations, significantly reduces vulnerability. Other vulnerable algoritms exist.
Backtracking Algorithmic Complexity Attacks Against a NIDS Thank you.