Decompression-Free Inspection: DPI for Shared Dictionary Compression over HTTP Anat Bremler-Barr Interdisciplinary Center Herzliya Shimrit Tzur David Interdisciplinary Center Herzliya & The Hebrew University, Jerusalem David Hay The Hebrew University, Jerusalem Yaron Koral Tel Aviv University 1

Outline Motivation Background ◦AC algorithm Our solution ◦The offline Phase ◦The online phase Experimental Results 2

Deep Packet Inspection (DPI) Search for patterns in the packets` payload Signatures-based NIDS ◦Intrusion Preventions Web-Application Firewalls ◦Leakage prevention ◦Content Filtering Challenges: ◦Thousands of known malicious patterns ◦Real time, link rate Security tools performance is dominated by the pattern matching engine (Fisk & Varghese 2002) 3

Compressed HTTP 4 19% increase in 8 month! 84.1% of the top 1,000 sites compress their traffic. Data compression is done by adding references to repeated data. There are two types of compression: ◦Intra-response compression – the references point to bytes within the response (Gzip/Deflate) ◦Inter-responses/connections compression – the references point to bytes in a separate file, called dictionary (Google’s SDCH).

Example – Intra-Response Compression File1.html: abcdefgabcd File2.html abcdxyzbcdtr Encode repeated strings by pointer: {distance, length} 5 TCP Connection Setup GET File1.html abcdefg(7,4) GET File2.html abcdxyz(6,3)tr

Example – Inter-Response Compression Dictionary: abcd File1.html: abcdefgabcd File2.html abcdxyzbcdtr Copy repeated strings from the dictionary: (address, length) 6 TCP Connection Setup GET File1.html Delta file: (0,4)efg(0,4) GET File2.html Delta file:(0,4)xyz(1,3)tr GET dictionary abcd

Current NIDS Operation (1) 7 ServerClient Http uncompressed NIDS GET \index.html Accept-Encoding: SDCH Scan for Intrusions Http uncompressed GET \index.html Accept-Encoding: SDCH

Current NIDS Operation (2) 8 ServerClient Http compressed NIDS GET \index.html Accept-Encoding: SDCH Do Not Scan/ Decompress, Scan, Compress Http compressed GET \index.html Accept-Encoding: SDCH

9 ServerClient Http compressed NIDS GET \index.html Accept-Encoding: SDCH Scan directly with no decompression Http compressed GET \index.html Accept-Encoding: SDCH

Our Solution: Decompression-Free Scanning Focused on inter-response compression Our algorithm works in two phases ◦Offline phase - Scanning the dictionary ◦Online phase - Scanning the delta files Works at the rate of the compressed traffic ◦Gain 56% improvement compared with scanning the plain-text directly 10

Outline Motivation Background ◦Aho-Corasick (AC) algorithm Our solution ◦The offline Phase ◦The online phase Experimental Results 11

Aho-Corasick (AC) Algorithm Finite State Machine (FSM) ◦Regular states, accepting states Goto function (black arrows) ◦g(state,symbol)  state Each state corresponds to a label- the sequence of characters on its goto path from the root. ◦The length of the label is the depth of the state Failure function (red arrows) ◦f(state)  state ◦Taken when there is no goto function ◦Goes to a state that its label is the longest suffix of the current state’s label s0s0 s7s7 s 12 s1s1 s2s2 s3s3 s5s5 s4s4 s 14 s 13 s6s6 s8s8 s9s9 s 10 s 11 C C E D B E D D B C A B A A The label of S 14 is BCAA g(S 11,B) = S 12 g(S 11,A) = ? Patterns: E BE BD BCAA BCD CDBCAB f(S 11 ) = S 13  g(S 11,A)  g(S 13,A)=S 14

Aho-Corasick Insights The automaton remembers only its current state ◦The input text ends with the label of current state ◦This label is the longest suffix in the text that can be a prefix of a match No future pattern can begin before this label s0s0 s7s7 s 12 s1s1 s2s2 s3s3 s5s5 s4s4 s 14 s 13 s6s6 s8s8 s9s9 s 10 s 11 C C E D B E D D B C A B A A

Outlines Motivation Background ◦Aho-Corasick (AC) algorithm Our solution ◦The offline Phase ◦The online phase Experimental Results 14

Accelerator Algorithm Idea The algorithm operates in two phases: The Offline Phase: ◦Scan the dictionary and store information about the pattern matching results The Online Phase: ◦Scan the delta file and skip almost all referenced bytes that were already scanned for patterns. 15

The Offline Phase The dictionary is scanned using AC (from its first byte and from s 0 ). We save the state after each byte CBACBDCAAEBD S5S5 S 12 S 11 S 10 S9S9 S8S8 S7S7 S0S0 S0S0 S3S3 S2S2 S0S0 s0s0 s7s7 s 12 s1s1 s2s2 s3s3 s5s5 s4s4 s14s14 s13s13 s6s6 s8s8 s9s9 s 10 s 11 C C E D B E D D B C A B A A State: We also save information of matched patterns that are found in the dictionary

Challenges Dictionary: Delta file: ABDB(5,4)AAB(1,4) The uncompressed data is: We copy from arbitrary position in the dictionary when the automaton in an arbitrary state ◦We show that no matter in what state and which symbol we start to copy, the resulting state is reachable via failure transitions from the saved state. 17 A B D B C D B C A A B B E A A Patterns/ Signatures: E BE BD BCAA BCD CDBCAB Types of matches: Right boundary Internal Left boundary DBEAACDBCABC

The Online Phase Scan the delta file: Uncompressed bytes - scan using AC. Copy instruction (p,x) ◦The compressed data that we already scanned in the offline phase. ◦We will save the scan for almost all these bytes. The internal match is trivial, see paper for details. 18

The Online Phase - Right Boundary When encountering copy instruction (p,x), We want to stop scanning and jump to state[p+x-1] ◦If the label of the state is longer than the copy- value  The label begins before the copy value  The context of this state is not as in the online scan  We take failure transitions to find state with sufficiently short label. ◦Otherwise  The label of the state is contained in the copy value  This is the longest suffix that can lead to a match 19

Example – Right Boundary Uncompressed data: …B 20 s0s0 s7s7 s 12 s1s1 s2s2 s3s3 s5s5 s4s4 s 14 s 13 s6s6 s8s8 s9s9 s 10 s 11 C C E D B E D D B C A B A A CBACBDCAAEBD S5S5 S 12 S 11 S 10 S9S9 S8S8 S7S7 S0S0 S0S0 S3S3 S2S2 S0S0 State: BCABCOPY(7,4): Go to State[10]=s 12. depth(s 12 ) > 4. Go to f(s 12 )=s 2 depth(s 2 ) ≤ 4 Current state is S 2

The Online Phase – Left Boundary When encountering copy instruction (p,x), We want to stop scanning and jump to state[p+x-1] ◦If the number of bytes we read from the copy value is less than the depth of the current state  The label of the state begins before the copied bytes  We scan the copy value till we reach a state that its label is shorter than the number of read bytes. ◦Otherwise  The label of the state is contained in the copy value  Both offline and online scans have the same context 21

Example – Left Boundary Uncompressed data: …B 22 s0s0 s7s7 s 12 s1s1 s2s2 s3s3 s5s5 s4s4 s14s14 s13s13 s6s6 s8s8 s9s9 s 10 s 11 C C E D B E D D B C A B A A CBACBDCAAEBD S5S5 S 12 S 11 S 10 S9S9 S8S8 S7S7 S0S0 S0S0 S3S3 S2S2 S0S0 State: CDBCCOPY(5,4): j=0 depth=1 Continue j=1 Depth=2 Continue j=2 Depth=3 Continue j=3 Stop scanning (depth(s 9 )≤3)

Outline Motivation Background ◦Aho-Corasick (AC) algorithm Our solution ◦The offline Phase ◦The online phase Experimental Results 23

Experimental Results Input: ◦google.com dictionary ◦Pages for 1000 most popular Google queries. Patterns ◦Snort The synthetic case ◦A patterns file for each input file so the input file has a different percentage of matches, from 25% to 100%. 24

The Algorithm Overheads 1. Traversing the failure transitions ◦In the right boundary 2. Scanning the copy value ◦In the left boundary 3. Memory consumption: ◦The additional information of the offline phase. ◦Total: 420 KB (per dictionary)  Can be further reduced by a variable-length pointer encoding. 25

Failure Transitions – Right Boundaries If length ≥ depth, no failure transition is taken In our experiments: ◦The average is 2.35 failure transitions per file  (average of 557 copy instructions per file) 26

Scanning the Copy Value - Left Boundary Compression ratio – compressed/uncompressed Scan ratio – scanned/uncompressed. Snort ◦low percentage of matches scan-ratio ~ compression ratio The synthetic case ◦high percentage of matches ◦Unrealistic case ◦scan-ratio is between 1.05 to 1.2 times compression- ratio. 27

Regular Expression Results Strings were extracted from the regular expression and were added to the pattern set. When needed, we use off-the-shelf perl compatible regular expression engine to scan additional parts of the text. The overhead of the regular expression is around 1% which is almost negligible 28

Questions?? 29

Regular Expression Very common in security purpose patterns. ◦In Snort, 55% of the rules contain regular expression. Composed of anchors and pcre tokens. For example, in the pattern: abc[1-9]*xyza{3,7} The anchors are: ◦abc ◦xyz The pcre tokens are: ◦[1-9]* ◦a{3,7} 30

Dealing with Regular Expression 1. The anchors are extracted from the regular expression offline. 2. The anchors are added to the patterns set. 3. If there is a regular expression which all its anchors were matched: ◦run an off the-shelf regular expression engine until, either a mismatch, a full pattern match, or the whole (limited) text is searched. 31

Regular Expression – Limited Search In most cases, we can limit the search in at least one direction. ◦If before the first anchor all tokens have a limited size, there is a bounded number of characters we should examine before the matched anchor. ◦If after the last anchor all tokens have a limited size there is a bounded number of characters we should examine after the matched anchor. 32

Memory Consumption 1. Doubling the size of the dictionary (for saving the offline scan results, one pointer per symbol) 2. Saving the matched list (for internal matches) Our experiments: ◦Match list size 40,000 ◦Dictionary size 116K symbols ◦Pointer size 17 bits Total memory consumption is 420 KB (per dictionary) ◦Can be further reduced by a variable-length pointer encoding. 33