accountable systems or how to catch a liar? jinyang li (with slides from authors of sundr and...
TRANSCRIPT
Accountable systems or
how to catch a liar?
Jinyang Li(with slides from authors of SUNDR and PeerReview)
What we have learnt so far
• Use BFT replication to handle (a few) malicious servers
• Strengths: – Transparent (service goes on as if no faults)– Minimal client modifications
• Weaknesses:– Breaks down completely if >1/3 servers are bad
Hold (bad) servers accountable• What is possible when >1/3 servers fail?
– E.g. 1 out of 1 servers fail?
• Let us consider a bad file server – Can it delete a client’s data?– Can it leak the content of a client’s data?– Can it change a client’s data to arbitrary values?– Can it show client A garbage and claim it’s from B?– Can it show client A old data of client B?– Can it show A (op1, op2) and B (op2, op1)?
Hold (bad) servers accountable
• Lessons:– Cannot prevent a bad server from
misbehaving at least once– We can do many checks to detect past
misbehaviors!
• Useful?
QuickTime™ and a decompressor
are needed to see this picture.
Fool me once, shame on you; fool me twice, shame on …
Case study I: SUNDR
• What’s SUNDR?– A (single server) network file system– Handle potential Byzantine server behavior– Can run on an un-trusted server
• Useful properties:– Tamper-evident
• Unauthorized operations will be immediately detected
– Can detect past misbehavior• If server drops operations, can be caught eventually
Ideal File system semantics
• Represent FS calls as fetch/modify operations– Fetch: client downloads new data– Modify: client makes new change visible to others
• Ideal (sequential) consistency: – A fetch reflects the sequence of modifications that
happen before it– Impossible when the server is malicious
• A is only aware of B’s latest modification via the server
• Goal: – get as close to ideal consistency as possible
Strawman File System
AModify f2 sig3
AModify f1 sig1
BFetch f4 sig2
AModify f2 sig3
BFetch f2 sig4
AModify f1 sig1
BFetch f4 sig2
AModify f1 sig1
BFetch f4 sig2
AModify f1 sig1
BFetch f4 sig2
sig1
sig2
AModify f2 sig3
sig3
File server
A: echo “A was here” >> /share/aaa
B: cat /share/aaa
userA
userB
Log specifies the total order
AModify f1 sig1
BFetch f4 sig2
AModify f2 sig3
BFetch f2 sig4
AModify f1 sig1
BFetch f4 sig2
AModify f2 sig3 The total order:
LogA ≤ LogB iff LogA is prefix of LogB
A’s latest log:
B’s latest log:
LogA
LogB
Detecting attacks by the server
AModify f2 sig3
AModify f1 sig1
BFetch f4 sig2
BFetch f2 sig3
A: echo “A was here” >> /share/aaa AModify f1 sig1
BFetch f4 sig2
AModify f1 sig1
BFetch f4 sig2
AModify f1 sig1
BFetch f4 sig2
AModify f2 sig3
A
BB: cat /share/aaa(stale result!)
File server
AModify f1 sig1
BFetch f4 sig2
BFetch f2 sig3b
AModify f1 sig1
BFetch f4 sig2
AModify f2 sig3a
A’s log and B’s log canno longer be ordered:LogA ≤ LogB, LogB ≤ LogA
A’s latest log:
B’s latest log:
Detecting attacks by the server
LogA
LogB
sig1
sig2
sig3a
What Strawman has achieved
High overhead, no concurrency Tamper-evident
• A bad server can’t make up ops users didn’t do
Achieves fork consistency• A bad server can conceal users’ ops from each
other, but such misbehavior could be detected later
Fork consistency is useful
• Best possible alternative to sequential consistency– If server lies only once, it has to lie forever
• Enable detections of misbehavior:– users periodically gossip to check violations– or deploy a trusted online “timestamp” box
SUNDR’s tricks to make strawman practical
1. Store FS as a hash tree– No need to reconstruct FS image from log
2. Use version vector to totally order ops– No need to fetch entire log to check for misbehavior
Trick #1: Hash tree
h0
h1
h3
h2
h4
h6
h5
h7
h9
h8
h10 h11 h12
D0
D1 D2D3
D4 D5 D6 D7 D8 D9 D10 D11 D12
Key property:h0 verifies the entire tree of data
Trick #2: version vector
• Each client keeps track of its version #• Server stores the (freshest) version vector
– Version vector orders all operations• E.g. (0, 1, 2) ≤ (1, 1, 2)
• A client remembers the latest version vector given to it by the server– If new vector does not succeed old one, detect
order violation!• E.g. (0,1,2) ≤ (0,2,1)
SUNDR architecture
• block server stores blocks retrievable by content-hash• consistency server orders all events
Untrusted Network
Untrusted Network
consistency server
block server
SUNDR server-side
SUNDRclient
SUNDRclient
userA
userB
SUNDR data structures: hash tree
• Each file is writable by one user or group• Each user/group manages its own pool of inodes
– A pool of inodes is represented as a hash tree
Hash files
• Blocks are stored and indexed by hash on the block server
data1
Metadata
H(data1)
H(data2)
H(iblk1)
data2
data3
data4H(data3)
H(data4)
iblk1
20-byte File Handle
i-node
Hash a pool of i-nodes
• Hash all files writable by each user/group
• From this digest, client can retrieve and verify any block of any file
2 20-byte
3 20-byte
4 20-byte
i-table i-node 2
i-node 3
i-node 4
20-byte digest
i-num
SUNDR FS
2 20-byte3 20-byte4 20-byte
digest
2 20-byte3 20-byte4 20-byte
digest
Superuser:
UserA:
SUNDR State
How to fetch “/share/aaa”?
/:Dir entry: (share, Superuser, 3)
Lookup “/”
/share:Dir entry: (aaa, UserA, 4)
Lookup “/share”
Fetch “/share/aaa”
digestUserB: …
234
digestGroupG:
SUNDR data structures:version vector
• Server orders users’ fetch/modify ops• Client signs version vector along with digest• Version vectors will expose ordering failures
Version structure
• Each user has its own version structure (VST)• Consistency server keeps latest VSTs of all users• Clients fetch all other users’ VSTs from server before
each operation and cache them• We order VSTA ≤ VSTB iff all the version numbers in VSTA
are less than or equal in VSTB
VSTA
Signature A
ADigest A
A - 1B - 1G - 1
VSTB≤ Signature B
BDigest B
A - 1B - 2G - 2
Update VST: An example
Consistency Server
B
AA-0
B-0
A: echo “A was here” >> /share/aaa
B: cat /share/aaa
DigA
A
A-1
B-1DigA
A
A-1
B-1DigA
A
A-1
B-1DigA
A
A-0
B-1DigB
B
A-1
B-2DigB
B
A-1
B-2DigB
B
A-0
B-1DigB
B
VSTA ≤ VSTB
Detect attacks
Consistency Server
BA: echo “A was here” >> /share/aaa
B: cat /share/aaa (stale!)
A
A-1
B-1DigA
A
A-0
B-0DigA
A
A-0
B-1DigA
B
A-1
B-1DigA
A A-0
B-1DigB
B
A-0
B-2DigB
B
A-0
B-2DigB
B
A’s latest VST and B’s can no longer be ordered:VSTA ≤ VSTB, VSTB ≤ VSTA
≤A-0
B-0DigA
A
Support concurrent operations
• Clients may issue operations concurrently• How does second client know what vector to sign?
– If operations don’t conflict, can just include first user’s forthcoming version number in VST
– But how do you know if concurrent operations conflict?
• Solution: Pre-declare operations in signed updates– Server returns latest VSTs and all pending updates,
thereby ordering them before current operation– User computes new VST including pending updates– User signs and commits new VST
Concurrent update of VSTs
Consistency ServerA
B
update: B-2
update: A-1
A: echo “A was here” >>/share/aaa
B: cat /share/bbb
A-0
B-0DigA
A
A-1
A-0
B-1DigB
B
A-0
B-1DigB
B
A-1 B-2
A-0
B-0DigA
A
VSTA ≤ VSTB
A-1
B-1DigA
A
A-1 A-1
B-1DigA
A
A-1
A-1
B-2DigB
B
A-1B-2A-1
B-2DigB
B
A-1B-2
SUNDR is practical
0
2
4
6
8
10
12
Create (1K) Read (1K) Unlink
NFSv2 NFSv3 SUNDR SUNDR/NVRAM
Seconds
Motivations for PeerReview
• Large distributed systems consist of many nodes
• Some nodes become Byzantine– Software compromise– Malicious/careless administrator
• Goal: detect past misbehavior of nodes– Apply to more general apps than FS
Challenges of general fault detection
• How to detect faults?• How to convince others that a node is (not)
faulty?
Overall approach• Fault := Node deviates from expected behavior• Obtain a signature for every action from each node• If a node misbehaves, the signature works a proof-of-misbehavior against
the node
Can we detect all faults?
• No– e.g. Faults affecting a node's
internal state
• Detect observable faults– E.g. bad nodes send a message
that correct nodes would not send
AA
X
CC
100101011000101101011100100100
0
Can we always get a proof?
• No – A said it sent X
B said A didn‘t
C: did A send X?
• Generate verifiable evidence:– a proof of misbehavior (A send wrong X)– a challenge (C asks A to send X again)
• Nodes not answering challenges are suspects
AA
X
BB
CC
?
I sent X!
I neverreceived
X!?!
• Treat each node as a deterministic state machine• Nodes sign every output message• A witness checks that another node outputs correct
messages– using a reference implementation and signed inputs
PeerReview Overview
M
PeerReview architecture• All nodes keep a log of
their inputs & outputs– Including all messages
• Each node has a set of witnesses, who audit its log periodically
• If the witnesses detect misbehavior, they– generate evidence
– make the evidence avai-lable to other nodes
• Other nodes check evi-dence, report fault
A's log
B's log
AA
BB
M
CCDD
EE
A's witnesses
M
PeerReview detects tampering
A B
Message
Hash chain
Send(X)
Recv(Y)
Send(Z)
Recv(M)
H0
H1
H2
H3
H4
B's log
ACK • What if a node modifies its log entries?
• Log entries form a hash chain• Signed hash is included with
every message Node commits to having received all prior messages
Hash(log)
Hash(log)
PeerReview detects inconsistency• What if a node
– keeps multiple logs?– forks its log?
• Witness checks if signed hashes form a single chain
H3'
Read X
H4'
Not found
Read Z
OK
Create X
H0
H1
H2
H3
H4
OK
"View #1""View #2"
Module B
PeerReview detects faults• Witness audits a node:
– Replay inputs on a reference implementation of the state machine
– Check outputs against the log
Module A
Module B
=?
LogNetwork
Input
Output
State machine
if ≠
Module A
PeerReview‘s guarantees
• Faults will be detected (eventually)– If a node if faulty:
• Its witness has a proof of misbehavior• Its witness generates a challenge that it cannot answer
– If a node is correct• No proof of misbehavor
• It can answer any challenge
PeerReview applications
• App #1: NFS server• Tampering with files
• Lost updates
• App #2: Overlay multicast• Freeloading
• Tampering with content
• App #3: P2P email• Denial of service
• Dropping emails
• Metadata corruption
• Incorrect access control
PeerReview’s performance penalty
• Cost increases w/ # of witnesses per node W
Baseline 1 2 3 4 5
100
80
60
40
20
0Avg traffic (Kbps/node)
Number of witnesses
Baseline traffic
What have we learnt?• Put constraints on what faulty servers can do
– Clients sign data, bad SUNDR server cannot fake data
– Clients sign version vector, bad server cannot hide past inconsistency
• Fault detection– Need proof of misbehavior (by signing actions)– Use challenges to distinguish slow nodes apart
from bad ones