HAIL (High-Availability and Integrity Layer) for Cloud Storage

HAIL (High-Availability and Integrity Layer) for Cloud Storage Kevin Bowers and Alina Oprea RSA Laboratories Joint work with Ari Juels

Cloud storage Cloud Storage Provider Storage server Web server Pros: Cons: Lower cost Easier management Enables sharing and Loss of control No guarantees of data access from anywhere availability Provider failures Client 2

Provider failures Amazon S3 systems failure downs Web 2.0 sites Twitterers lose their faces, others just want their data back Computer World, July 21, 2008 Customers Shrug Off S3 Service Failure At about 7:30 EST this morning, S3, Amazon.com’s online storage service, went down. The 2-hour service failure affected customers worldwide. Wired, Feb. 15, 2008 Temporary unavailability Spectacular Data Loss Drowns Sidekick Users October 10, 2009 Loss of customer data spurs closure of online storage service 'The Linkup‘ Network World, Nov 8, 2008 Permanent data loss How do we increase users’ confidence in the cloud? 3

Outline Proofs of Retrievability – Constructions and practical aspects – Limitations HAIL goals and adversarial model HAIL protocol design – – – – Encoding layer Decoding layer Challenge-response protocol Redistribution of shares in case of failures HAIL parameter tradeoffs Open problems 4

PORs: Proofs of Retrievability Client outsources a file F to a remote storage provider Client would like to ensure that her file F is retrievable The simple approach: client periodically downloads F This is resource-intensive! What about spot-checking instead? – Sample a few file blocks periodically – If file is not stored locally, need verification mechanism (e.g., MACs for each file block) 5

Spot-checking Cloud Storage Provider F B1 B4 B7 T1 T2 T3 MACk[B4] T1 T2 T3 Client k 6

Spot-checking Cloud Storage Provider F B1 B4 B7 T1 T2 T3 Small corruptions go undetected Client k 7

Error correcting code Cloud Storage Provider Corrects small corruption F Parity blocks Client k 8

ECC MAC Cloud Storage Provider F B1 B4 B7 T1 T2 T3 T4 P1 Parity blocks MACs over file and parity blocks Detect large corruption through spot checking Corrects small corruption through ECC Client k 9

Query aggregation Cloud Storage Provider F Parity blocks Challenge Client MACs over aggregation of blocks Response k 10

Practical considerations Applying such an ECC to all of F is impractical Instead, we can stripe the ECC If adversary knows the stripe structure, she can corrupt selectively 11

Selective corruption Adversary targets a particular stripe File can not be recovered The probability that the client detects the corruption through sampling is small if stripes are small Practical code parameters encode hundreds of bytes at a time (e.g., Reed-Solomon (255, 223, 32)) 12

Adversarial codes: hide ECC stripes Do secret, randomized partitioning of F into stripes – E.g. use secret key to generate pseudorandom permutation and then choose stripes sequentially Encrypt and permute parity blocks The encoding is still systematic But adversary does not know where stripes are, so adversary cannot feasibly target a stripe! 13

POR papers Proofs of Retrievability (PORs) – Juels-Kaliski 2007 Proofs of Data Possession (PDPs) – Burns et al. 2007 – Erway et al. 2009 Unlimited queries using homomorphic MACs – Shacham-Waters 2008 – Ateniese, Kamara and Katz 2009 Fully general query aggregation in PORs – Bowers, Juels and Oprea 2009 – Dodis, Vadhan and Wichs 2009 14

When PORs fail Cloud Storage Provider FF decoder Challenge Response Unrecoverable Client k 15

Outline Proofs of Retrievability – Constructions and practical aspects – Limitations HAIL goals and adversarial model HAIL protocol design – – – – Encoding layer Decoding layer Challenge-response protocol Redistribution of shares in case of failures HAIL parameter tradeoffs Open problems 16

HAIL goals Resilience against cloud provider failure and temporary unavailability Use multiple cloud providers to construct a reliable cloud storage service out of unreliable components – RAID (Reliable Array of Inexpensive Disks) for cloud storage under adversarial model Provide clients or third party auditing capabilities – Efficient proofs of file availability by interacting with cloud providers 17

RAID (Redundant Array of Inexpensive Disks) Stripe Data block Data block Data block Parity block B1 B X B3 P1 B1 B2 B3 2 B1 B3 P1 Shift from monolithic, high-performance drives to cheaper drives with redundancy 18

RAID in the Cloud Provider A Provider B Provider C Provider D Fuse together cheap cloud providers to provide highquality (reliable) abstraction – E.g., Memopal offers 0.02 / GB / Month storage on a 5-year contract vs. Amazon at 0.15 / GB / Month 19

But the cloud is adversarial! Provider A Provider B Provider C Provider D RAID designed for benign failures (drive crashes) Static adversaries are not realistic A mobile adversary moves from provider to provider – System failures and corruptions over time – Corrupts a threshold of providers in each epoch (b out of n) 20

Mobile adversary Provider A Provider B Provider C Provider D Combination of proactive and reactive models – Separate each server into code base and storage base Code base of servers cleaned at beginning of epoch (e.g., through reboot) At most b out of n server have corrupted code in each epoch – Challenge-responses used for detection of failure Corrupted storage recovered when failure is detected 21

HAIL protocols File encoding – Distribute a file across n storage providers – Add redundancy to tolerate provider failures – Small state stored locally by client (including secret key) File decoding – Recover original file by contacting a threshold of providers – Tolerate provider failures or unavailability Challenge-response protocol – Executed a number of times per epoch – Enables clients to perform integrity checks by contacting a threshold of providers – Detects failures early and enhances data availability Share redistribution – When failure detected, clients reconstruct shares from redundancy encoded in other providers 22

Outline Proofs of Retrievability – Constructions and practical aspects – Limitations HAIL goals and adversarial model HAIL protocol design – – – – Encoding layer Decoding layer Challenge-response protocol Redistribution of shares in case of failures HAIL parameter tradeoffs Open problems 23

First idea: file replication with POR Provider A F Provider B Provider C F F Parity MACs Parity MACs Parity MACs POR Response POR Challenge POR Challenge POR Response POR Challenge POR Response F Client 24

File replication with POR: Issues Provider A FF Provider C F F Parity MACs Provider B Parity MACs Parity MACs Compute different MACs per provider Large encoding overhead Large storage overhead due to replication Client 25

Use redundancy across servers Provider A Provider B Provider C F F F Block i Block i Fi Client Block i F Fi Fi Sample and check consistency across providers 26

Small-corruption attack Provider A Fi F Provider B Fi F Provider C Fi F File can not be recovered after [n/b] epochs Client The probability that client samples the corrupted block is low 27

Replication with server code Provider A Provider B F F F Parity Provider C Parity Parity Still vulnerable to small-corruption attack, once corruption exceeds the error correction rate of server code Large storage overhead due to replication Client 28

Dispersal erasure code Primary servers (k) PA Secondary servers (n-k) PB PC PD PE 128 bit F1 F2 Original file F F3 Stripe Dispersal code parity File can be recovered from any k available servers For encoding efficiency, use striping for 128-bit blocks F 29

Two encoding layers PA PB F1 PC F2 PD F3 PE Dispersal code parity Server code Dispersal code reduces storage overhead of replication with similar availability guarantees Server code improves resilience to small-corruption attack 30

Checking for correct encoding PA PB Client PC PD PE Check that stripe is a codeword in dispersal code 31

Aggregation of stripes PA PB PC PD PE 1 α α2 Client Check that linear combination of stripes is a codeword 32

Comparison File replication with POR F F Parity MACs F Parity MACs Parity MACs HAIL:Two encoding layers (dispersal and server code) - Large storage overhead due to replication - Redundant MACs for POR - Large encoding overhead - Verifiable by client only Increased lifetime Optimal storage overhead for given availability level Uses cross-server redundancy for verifying responses Reasonable encoding overhead Public verifiability 33 - Limited lifetime

Increase protocol lifetime PA F1 PB F2 PC PD PE F3 MAC Authenticate stripes with MACs One MAC per block - Large storage overhead - How can the MACs from multiple stripes be aggregated? 34

Integrity-protected dispersal code PA F1 PB F2 PC PD F3 PE PRFk1(pos) Embed integrity information into parity blocks of dispersal code Can check linear combination of MACs knowing only linear combination of blocks 35

HAIL protocols Encoding – Two layers of error correction: dispersal code and server code – Integrity-protected dispersal code used to reduce storage overhead – Server code is adversarial erasure code Decoding – Reverse of encoding, using two layers of error correction Tradeoffs: – Erasure dispersal code: tolerates n-m-1 failures per round, but decoding requires brute force in case of errors (do not know the positions of erasures) – Error-correcting dispersal code: tolerates up to b (n-m-1)/2 failures per round 36

HAIL protocols, cont Challenge-response – – – – Executed in each time round a number of times Challenge: a number of row positions Response: aggregated row Verification: response should be a codeword in dispersal code and composite MAC should be valid Redistribution of shares: – Invoked when corruption of a fragment is detected by challengeresponse – Reconstruction done by client and involves downloading m correct file fragments 37

HAIL availability 38

Frequency of challenges 39

Encoding Performance HAIL requires two levels of encoding Order is important! 40

Encoding Security Security of the MAC depends on the size of the finite field used to perform Reed-Solomon encoding. Most Reed-Solomon codes are implemented over bytes, or at most 4-byte words (typical integer representation) 32-bit security is low from a cryptographic viewpoint Operating over larger symbols is slow – Larger encodings can be generated by combining several smaller encodings – Or, they can be implemented using extension fields To speed up larger symbol encoding, need fast operations in large Galois Fields – Work with Jianqiang Luo and Lihao Xu at Wayne State Univ. 41

Encoding Throughput Improvement 42

Decoding Throughput Improvement 43

Accelerated Encoding Throughput 44

Accelerated Decoding Throughput 45

Effect of Placement on Throughput 46

Summary HAIL is an extension of RAID into the cloud High availability and tolerance to adversarial failures – Low storage overhead due to integrity-protected dispersal code Enables client-side integrity checks – Low bandwidth for challenge-response due to aggregation Papers: – K. Bowers, A. Juels, and A. Oprea. Proofs of Retrievability: Theory and Implementation. ACM CCSW ’09. – K. Bowers, A. Juels, and A. Oprea. HAIL: High Availability and Integrity Layer for Cloud Storage. ACM CCS ’09. http://www.rsalabs.com/ 47