This document summarizes research on breaking encryption schemes that use the RC4 stream cipher. The researchers discovered new biases in the RC4 keystream that allow recovering plaintext more efficiently. They applied these biases to break WPA-TKIP encryption and decrypt HTTPS cookies. For WPA-TKIP, simulating traffic captures allowed decrypting packets and the message integrity check key within an hour. For HTTPS, biases were combined to decrypt 16-character cookies in around 75 hours by manipulating requests sent to a target site. The work significantly advances attacks against protocols still relying on the weakened RC4 cipher.

1. All Your Biases Belong To Us:
Breaking RC4 in WPA-TKIP and TLS
Mathy Vanhoef and Frank Piessens, KU Leuven
USENIX Security 2015

2. RC4
2
Intriguingly simple stream cipher
WEP
WPA-TKIP
SSL / TLS PPP/MPPE
And others ...

3. RC4
3
Plaintext CiphertextKeystreamRC4
Key
Intriguingly simple stream cipher

4. Is RC4 still used?!
4
ICSI Notary: TLS connections using RC4
50%
30%
13%
0%
10%
20%
30%
40%
50%
60%
March 2013 Februari 2015 July 2015
RC4 fallback not taken into account!

5. RC4 Fallback
5
Client Server
ClientHello: without RC4 Browser first tries without RC4
ServerHello: use AES

6. Alert: Handshake Failed
RC4 Fallback
6
ClientHello: without RC4 Browser first tries without RC4
If that fails …
Client Server

7. ClientHello: with RC4
ServerHello: use RC4
RC4 Fallback
7
Client Server
Alert: Handshake Failed
ClientHello: without RC4 Browser first tries without RC4
If that fails …
… fallback to RC4

8. ClientHello: with RC4
ServerHello: use RC4
RC4 Fallback
8
Client Server
Alert: Handshake Failed
ClientHello: without RC4 Browser first tries without RC4
Forgeable by attacker!
… fallback to RC4
 13% estimate is a lower bound
 Force connection (which we
assumed secure) to use RC4

9. Our Goal: further kill RC4
9
New Biases Plaintext Recovery
Break WPA-TKIP Attack HTTPS

10. First: Existing Biases
10
Distribution keystream byte 2
Pr 𝒁 𝟐 = 𝟎 =
𝟐
𝟐𝟓𝟔
[MS01]

11. First: Existing Biases
11
Distribution keystream byte 1 (to 256)

12. First: Existing Biases
12
Distribution keystream byte 1 (to 256)
AlFardan et al. ‘13:
first 256 bytes biased
Short-term biases

13. Long-Term Biases
13
A B S A B
Fluhrer-McGrew (2000):
 Some consecutive values are biased
Examples: 0, 0 and (0, 1)
Mantin’s ABSAB Bias (2005):
 A byte pair (𝐴, 𝐵) likely reappears

14. Fluhrer-McGrew: only 8 out
of 65 536 pairs are biased
Search for new biases
14
Traditional emperical approach:
 Generate large amount of keystreams
 Manually inspect data or graph
How to automate
the search?

15. Search for new biases
15
Hypothesis tests!
 Uniformly distributed: Chi-squared test.
 Correlated: M-test (detect outliers = biases)
Traditional emperical approach:
 Generate large amount of keystreams
 Manually inspect data or graph
Allows a large-scale search,
revealing many new biases

16. Biases in Bytes 258-513
16
Example: keystream byte 258

17. Biases in Bytes 258-513
17
Example: keystream byte 320

18. Biases in Bytes 258-513
18
Example: keystream byte 352
Biases quickly
become quite weak

19. New Long-term Bias
19
(𝑍256∙𝑤, 𝑍256∙𝑤+2) = (128, 0)
with probability 2−16
(1 + 2−8
)
128 0 ...
Every block of 256 bytes

20. Additional Biases
20
See paper!

21. Our Goal: further kill RC4
21
New Biases Plaintext Recovery
Break WPA-TKIP Attack HTTPS

22. Existing Methods [AlFardan et al. ‘13]
22
Plaintext encrypted under
several keystreams
Ciphertext Distribution Plaintext guess 𝜇
Induced keystream
distribution
Verify guess: how close to
real keystream distribution?

23. Example: Decrypt byte 1
23
Ciphertext Distribution

24. Example: Decrypt byte 1
24
RC4 & Ciphertext distribution

25. Example: Decrypt byte 1
25
If plaintext byte 𝜇 = 0x28: RC4 & Induced
𝜇 = 0x28 has low likelihood

26. Example: Decrypt byte 1
26
If plaintext byte 𝜇 = 0x5C: RC4 & Induced
𝜇 = 0x5C has higher likelihood

27. Example: Decrypt byte 1
27
If plaintext byte 𝜇 = 0x5A: RC4 & Induced
𝜇 = 0x5A has highest likelihood!

28. Types of likelihood estimates
28
Previous works: pick value with highest likelihood.
Better idea: list of candidates in decreasing likelihood:
 Most likely one may not be correct!
 Prune bad candidates (e.g. bad CRC)
 Brute force cookies or passwords
How to calculate list of candidates?

29. 1st idea: Generate List of Candidatess
29
Gist of the Algorithm: Incremental approach
Calculate candidates of length 1, length 2, ...
1
2
𝑛
1
2
𝑛
1
2
𝑛
...

30. 2nd idea: abusing the ABSAB bias
30
Assume there’s surrounding known plaintext
 Derive values of A, B
 Combine with ABSAB bias to (probablisticly) predict A′, B′
 Ordinary likelihood calculation over only (A′, B′)
A B S A’ B’
Known Plaintext Unknown Plaintext
Likelihood estimate:
!

31. Our Goal: further kill RC4
31
New Biases Plaintext Recovery
Break WPA-TKIP Attack HTTPS

32. TKIP Background
32
How are packets sent/received?
1. Add Message Integrity Check (MIC)
2. Add CRC (leftover from WEP)
3. Add IV (increments every frame)
4. Encrypt using RC4 (per-packet key)
Encrypted
MICDataIV CRC

33. Flaw #1: TKIP Per-packet Key
33
Key-Mix
Key Sender MAC 𝐼𝑉
packet key
Anti-FMS(𝐼𝑉0, 𝐼𝑉1)
 𝐼𝑉-dependent biases in keystream
[Gupta/Paterson et al.]
Avoid weak keys which broke WEP

34. Flaw #2: MIC is invertible
34
If decrypted, reveals MIC key
MICDataIV CRC
 With the MIC key, an attacker can inject and
decrypt some packets [AsiaCCS ‘13]

35. Goal: decrypt data and MIC
35
If decrypted, reveals MIC key
MICDataIV CRC
Generate identical packets (otherwise MIC changes):
 Assume victim connects to server of attacker
 Retransmit identical TCP packet
 List of plaintext candidates (unknown MIC and CRC)
 Prune bad candidates based on CRC

36. Evaluation
36
Simulations with 230
candidates:
 Need ≈ 224
captures to decrypt with high success rates
Emperical tests:
 Server can inject 2 500 packets per second
 Roughly one hour to capture sufficient traffic
 Successfully decrypted packet & found MIC key!

37. Our Goal: further kill RC4
37
New Biases Plaintext Recovery
Break WPA-TKIP Attack HTTPS

38. TLS Background
38
Client Server
 Focus on record protocol with RC4 as cipher
Handshake protocol
Negotiate keys
Record protocol
Encrypt data

39. Targeting HTTPS Cookies
39
Previous attacks only used Fluhrer-McGrew (FM) biases
We combine FM bias with the ABSAB bias
Must surround cookie with known plaintext
1. Remove unknown plaintext arround cookie
2. Inject known plaintext arround cookie

40. Example: manipulated HTTP request
40
User-Agent: Mozilla/5.0 (Windows NT 6.1; WOW64;
Trident/7.0; rv:11.0) like Gecko
Host: a.site.com
Connection: Keep-Alive
Cache-Control: no-cache
Cookie: auth=????????????????; P=aaaaaaaaaaaaaaaaa
Surrounded by known
plaintext at both sides
Headers are
predictable

41. Preparation: manipulating cookies
41
Clienta.site.com fake.site.com
HTTPS insecure
Remove & inject
secure cookies!

42. Performing the attack!
42
JavaScript: Cross-Origin requests in WebWorkers

43. Performing the attack!
43
Keep-Alive connection to generate them fast

44. Performing the attack!
44
Combine Fluhrer-McGrew and ABSAB biases

45. Decrypting 16-character cookie
45
Takes 75 hours with 4450 requests / second
Ciphertext copies times 227

46. Decrypting 16-character cookie
46
DEMO!
rc4nomore.com

47. Questions?
May the bias be ever in your favor