1. ReplayConfusion: Detecting Cache-based Covert Channel Attacks Using Record and Replay
Mengjia Yan, Yasser Shalabi, Josep Torrellas
University of Illinois at Urbana-Champaign
Say uiuc
MICRO October 2016

2. Mengjia Yan | ReplayConfusion
Motivation
Cache-based covert channel attacks
Communicate through cache conflicts
Serious security threat
Ubiquitous attack scenario: cloud
Bypass security policy; no trace left
Existing solutions unable to detect all attacks
Contribution: ReplayConfusion
High-coverage detection mechanism
Trojan
(sender)
Spy
(receiver)
Core L1
Core L1
Core L1
Core L1
LLC Cache
Explain trojan better
Trojan process : malicious process inject into a victim process
has secret information and leak it to a spy process
It has malicious component, but appear to legal to users
Starts with the motivation is there is a type of attack: ….
Ebds with our contribution replayconfusion is a mechanism that provides high-coverage detect for this attack

3. Contribution: ReplayConfusion
Mengjia Yan | ReplayConfusion
Contribution: ReplayConfusion
Observations:
Trojan/Spy rely on specific mapping addressescaches
Attack follows a repeating pattern when transmitting
Effects:
Substantially disrupt cache miss pattern
Retain the repeating pattern
Change mapping of addressescaches
Re-mapping is different for each process ?
Cache
misses
Explain record and replay first, non-deterministic events into a log
Finish by Saying separate attackers from benign programs
RnR: as program executes, record ND into a log and obtain timeline from PMU
Cache miss rate timeline!!
record
cache
replay
(different
mapping)
Cache
misses
cache

4. Mengjia Yan | ReplayConfusion
Outline
Background
Attack Protocols
ReplayConfusion
Observations
Detection Framework
Detection Example
Summary

5. Cache-based Covert Channel Attack
Mengjia Yan | ReplayConfusion
Cache-based Covert Channel Attack
Basic cache organization:
Slice(i.e. Bank), Set, Way
Cache mapping function
Approach: Prime+Probe
Cache
Mapping
Function
Slice ID
Set ID
Tag
Physical
Address
Can be reverse engineered
trojan
spy
prime 1 1
Example: 2 groups
From now on, I use one access to represent multiple accesses to sets
Probe fills all cache sets, so it can be used as next prime stage
Slice 0
Slice 1
Set 0
probe 1 1
Set 1
Time

6. Taxonomy of Attack Protocols
Mengjia Yan | ReplayConfusion
Taxonomy of Attack Protocols
trojan
spy
Round-robin
Parallel
Round-Robin
Time
Parallel 1 1
Attack Protocols
Single Group
ReplayConfusion detects all the attacks
Space
Multiple Groups 1 1
The paper presents taxonomy of existing .. Through cache conflicts.
Time
Time
group 0
group 1
slice0
slice1
set0
set1
slice0
slice1
set0
set1
slice0
slice1
set0
set1

7. Mengjia Yan | ReplayConfusion
Observations
Observation 1:
Trojan/spy rely on a specific cache mapping function
Observation 2:
Attack follows a repeating pattern when transmitting
slice0
slice1
set0
set1
trojan
group 0 1
spy
group 1
slice0
slice1
set0
set1 1 1
slice0
slice1
set0
set1 1
Time

8. ReplayConfusion Detection Approach
Mengjia Yan | ReplayConfusion
ReplayConfusion Detection Approach
Observations:
Trojan/Spy rely on a specific cache mapping function
Attack follows a repeating pattern when transmitting
Effects:
Substantially disrupt cache miss pattern
Retain the repeating pattern
Change mapping of addressescaches
Re-mapping is different for each process ?
Cache
misses
Move boxes to fit text
record
cache
Replay
(different
mapping)
Cache
misses
cache

9. Replay Confusion Detection Approach
Mengjia Yan | ReplayConfusion
Replay Confusion Detection Approach ?
record
cache
Cache
misses
Cache
misses
Replay
cache
Record and Replay
Design new HW mapping addresses  caches
Analyze cache miss rate timelines
To implement ReplayConfusion, we need to support three mechanisms
Capo/a:/
Existing mature technique
e.g. Capo, Cyrus …
Requirements:
Small impact on benign programs
Big impact on attacks
Look for a repeating pattern in the timeline of the cache miss rate difference

10. Designing New Cache Mapping Functions
Mengjia Yan | ReplayConfusion
Designing New Cache Mapping Functions
Goal
Small impact on benign programs
Big impact on attacks
Set Index Function
Swap or flip bits within index field
Set Index
Tag
Block
Offset
Phys
Addr
Use the swap for later
Slice Selection Function
Replace the bits in the function with nearby ones
xor
Slice ID

11. Analyzing Cache Miss Rate Timelines
Mengjia Yan | ReplayConfusion
Analyzing Cache Miss Rate Timelines
Compute timeline of the difference in cache miss rates
(Recording miss rate timeline) – (Replay miss rate timeline)
Use auto-correlation* to detect repeating pattern in the timeline of the cache miss rate difference
Look for a fluctuating pattern in the auto-correlation
Benign programs
Attacks
Diff Value
Small values mostly
Large values when transmitting
Diff Pattern
No pattern
Repeating pattern
Compute the timeline difference of cache miss rate
*A statistical technique that discovers repeating patterns in a signal.

12. Mengjia Yan | ReplayConfusion
Detection Example
Experiment #1:
Bzip2
(co-run with h264ref)
Experiment #2:
Spy in attack
(co-run with Trojan)
Miss Rate Timeline in
Record
Replay
Miss Rate Difference
Timeline
X axis: time
Y axis: cache miss rate

13. Mengjia Yan | ReplayConfusion
Detection Example
Experiment #1:
Bzip2
(co-run with h264ref)
Experiment #2:
Spy in attack
(co-run with Trojan)
Miss Rate Timeline
Difference
Auto-correlation
Repeat miss rate difference timeline from last slide
Apply autocorrelation to the difference timeline
Auto correlation computes the similarity of a signal to itself with certain
peak -> high value, very similar to itself
Low value-> not similar

14. Mengjia Yan | ReplayConfusion
More in the Paper
Details on the taxonomy of cache-based covert channel attacks
More detection results
Attacks using different protocols
Attacks with background noise
Attacks with small group size
More benign programs
Detailed discussion about robustness of ReplayConfusion
Discussion of related works

15. Mengjia Yan | ReplayConfusion
Conclusion
Characteristics of cache-based covert channel attacks:
Trojan/spy communication is tuned to mapping of addresses to caches
Miss rate pattern repeats when transmitting bits
ReplayConfusion
Use RnR to execute the same program on machines with different mappings of addresses to caches in replay
Compute the timeline of the miss rate difference between record and replay
Detect repeating patterns  detect attack

16. ReplayConfusion: Detecting Cache-based Covert Channel Attacks Using Record and Replay
Mengjia Yan, Yasser Shalabi, Josep Torrellas
University of Illinois at Urbana-Champaign
Say uiuc
MICRO October 2016

17. Mengjia Yan | ReplayConfusion
Thank You

18. Mengjia Yan | ReplayConfusion
Backup Slides

19. Evaluation Result Benign Programs
Mengjia Yan | ReplayConfusion
Evaluation Result Benign Programs
h264ref
sjeng
gobmk
stream

20. Mengjia Yan | ReplayConfusion
Experiment Setup
System: Ubuntu 10.4 with 4GB memory
4 in-order core, 32KB private L1 cache, 2MB shared L2 cache
L2: 8-way associative, 4 slices, 64B/block

21. Mengjia Yan | ReplayConfusion
Example
(a) Cache miss rate timeline
(b) Cache miss rate autocorrelogram
(c) Cache miss rate difference timeline
(d) Cache miss rate difference autocorrelogram

22. Evaluation Result Attacks using parallel protocols
Mengjia Yan | ReplayConfusion
Evaluation Result Attacks using parallel protocols
2-group
1-group, ¼ cache, set-based
1-group, ¼ cache, slice-based
1-group, unaware

23. Related Work Defense Detection Cache Partition Add noise to timer
Mengjia Yan | ReplayConfusion
Related Work
Either not not applicable or too much overhead
Defense
Cache Partition
Add noise to timer
Detection
Hexpad: high cache access rate
Chiappetta et al. : correlation between sender and receiver
CC-Hunter: detect alternate pattern of conflicts
Not work effectively
Unable to detect advanced attacks
May have high false positives
Only effective to attacks using a specific type of protocols

24. Operations of ReplayConfusion
Mengjia Yan | ReplayConfusion
Operations of ReplayConfusion ⑤
Log
Cache Miss Rate Timeline ② ④ ② ④
RnR Module
Cache Configuration
Manager
Cache Profile
Manager SW ① ③
PMU HW F0
Fsel
Address
Mapping
Memory
Address Fn
Slice ID
Set
Index
Tag
Cache Address
Computation Unit