consumption with Popcorn Trinabh Gupta The University of Texas at - - PowerPoint PPT Presentation

▶
consumption with popcorn
SMART_READER_LITE
LIVE PREVIEW

consumption with Popcorn Trinabh Gupta The University of Texas at - - PowerPoint PPT Presentation

Mar 01, 2024 23 likes •277 views

Scalable and private media consumption with Popcorn Trinabh Gupta The University of Texas at Austin give me The Godfather 90 minutes/day The Godfather User media consumption has increased database of request trace, movie ratings , etc.

slide-1
SLIDE 1

Scalable and private media consumption with Popcorn

Trinabh Gupta

The University of Texas at Austin

slide-2
SLIDE 2

90 minutes/day

The Godfather give me The Godfather database of request

trace, movie ratings,

etc.

User media consumption has increased … … leading to large centralized datasets …

??

movie 1 movie 2 movie 3 movie 1 movie 2 movie 3 researcher anonymized dataset

  • f movie ratings

de-anonymized dataset

  • f movie ratings

… subject to risks such as server hacks, accidental disclosures, etc.

slide-3
SLIDE 3
slide-4
SLIDE 4

How can we build a Netflix-like system that a) provably hides media diet, b) has low dollar cost, and c) is compatible with commercial media

streaming?

slide-5
SLIDE 5

Private Information Retrieval (PIR) provably hides requests but …

wants The Godfather The Godfather [hidden] give me [hidden]

  • Each request must touch the entire library.
  • There is a tension between overhead and content protection.
  • PIR assumes fixed-size objects, but media sizes vary.
slide-6
SLIDE 6

Its per-request dollar cost is 3.87x times that of a non-private baseline. Popcorn tailors PIR for media to meet our three requirements.

6

slide-7
SLIDE 7

Rest of this talk

  • Background on PIR.
  • Challenges of using PIR (in detail).
  • Design (tailoring of PIR) and evaluation of Popcorn.

7

slide-8
SLIDE 8

Pick a subset of {1, 2, 3, 4, 5} randomly No collusion

Background on information-theoretic PIR (ITPIR)

Server1 Client M1 = Reply1 Reply2

Reply1 = M2 M4

M1 M2 M3 M4 M5

01111001….... 010111000…. 10101011…… 11100000…… 0011000.…….

M1 M2 M3 M4 M5

01111001….... 010111000…. 10101011…… 11100000…… 0011000.…….

Ex: {3, 4} Ex: {1, 2, 4, 5} Server 2 {2, 4}

Reply 2 = M1 M2 M4

wants M1

slide-9
SLIDE 9

Computational PIR (CPIR) from 10,000 feet

Client

M1 M2 M3 M4 M5

01111001….... 010111000…. 10101011…… 11100000…… 0011000.…….

  • one server
  • instead of XORs, expensive server-side

cryptographic operations

slide-10
SLIDE 10

Given these, how can we build a system that controls content and is low cost?

cheap operations (XORs) but process entire library per request CPIR expensive operations and process entire library per request ITPIR assumes fixed-size objects assumes fixed-size objects content can disseminate in an uncontrolled manner content disseminates in a controlled manner

Challenges of using PIR

slide-11
SLIDE 11

Server 1 (library owner) Client Server 2

ITPIR CPIR

Popcorn composes ITPIR and CPIR to get desirable properties from both

Enc(K1, M1) Enc(K2, M2) Enc(K3, M3) Enc(K4, M4) Enc(K5, M5) Enc(K1, M1) Enc(K2, M2) Enc(K3, M3) Enc(K4, M4) Enc(K5, M5)

different administrative domains Enc(K1, M1) K1

K1 K2 K3 K4 K5

Key library

encrypted movie key to decrypt movie

slide-12
SLIDE 12

cheap operations (XORs) but process entire library per request CPIR expensive operations and process entire library per request ITPIR assumes fixed-size objects assumes fixed-size objects content can disseminate in an uncontrolled manner content disseminates in a controlled manner

Popcorn

Challenges of using PIR

slide-13
SLIDE 13

Reply = M1 M3 M4 M5

Observation: Very similar disk I/O for each request! Benefits of batching:

  • Disk I/O transfers are amortized.
  • CPU cycles are reduced as matrix multiplication algorithms

exploit cache locality.

Popcorn batches requests to amortize the

  • verhead of ITPIR

{1, 3, 5} {1, 3, 4, 5} {2, 4}

  

Reply = M1 M3 M5

 

Reply = M2 M4

M1 M2 M3 M4 M5

01111001….... 010111000…. 10101011…… 11100000…… 0011000.…….

Server1 Pick a subset of {1, 2, 3, 4, 5} randomly Client 1 Client 2 Client 3

slide-14
SLIDE 14

time client A client B client C

Straw man: Group requests that arrive during an epoch

epoch client A wait for server to form batch start handling A, B, C client A’s playback buffer client perceived delay = epoch + epsilon first chunk

  • f movie

Client’s view:

slide-15
SLIDE 15

time client A client B client C

Straw man: Group requests that arrive during an epoch

epoch start handling A, B, C

Small batch, small delay Large batch, large delay

Issue: Hard to get both small delay and large batch

Server’s choices:

slide-16
SLIDE 16

t = times at which a client needs movie chunks

Popcorn exploits streaming to form large batches with small startup delay

t = 0 chunks of a movie t = t = 2 = time it takes to consume a single chunk

Observation: Client needs only the first chunk immediately.

t t t

t = 3t

slide-17
SLIDE 17

2nd library column 1st library column

0101111101101101001010010010010111001101111

Narrow first column => small startup delay

3rd library column

0101111 10110110100 1011010010010010111001101111 1011011 1001001 00100010011 11100011101

Movie 1 Movie 2 Movie 3

0000100011110001110100100100 0011111000000011010101010111 1001001 11100011101

Movie 4

0011111000000011010101010111

…

Wider columns => longer processing times …

… but bigger batches

Movie 1 …

slide-18
SLIDE 18

ITPIR CPIR Popcorn content can disseminate in an uncontrolled manner content disseminates in a controlled manner content disseminates in a controlled manner cheap operations but process entire library per request expensive operations, process entire library per request cheap operations, process entire library per batch assumes fixed-size

  • bjects

assumes fixed-size

  • bjects

?

slide-19
SLIDE 19

Popcorn exploits compression to address fixed-size requirement

Length Pad

  • Small variations in bitrate have limited impact on user satisfaction

[SIGCOMM 11, LANC 11, CCNC 12].

  • 85% of movies close to the average size.

Length of Oavg

slide-20
SLIDE 20

Outline

Background on PIR. Design (tailoring of PIR) of Popcorn.

  • Evaluation of Popcorn.

Popcorn

slide-21
SLIDE 21

Experiment method

Baselines:

  • Non-private system (Apache server)
  • State-of-the-art CPIR [XPIR PETS16]
  • State-of-the-art ITPIR [Percy++]
  • ITPIR++: ITPIR extended with the straw man batching scheme

Netflix-like library: 8000 movies, 90 minutes, 4Mbps Workload: 10K clients arrive within 90 minutes according to a Poisson process Estimate per-request dollar cost using Amazon’s pricing model

  • CPU: $0.0076/hour
  • Disk I/O bandwidth: $0.042/Gbps-hour
  • Network: $0.006/GB
slide-22
SLIDE 22

System # of CPUs Disk I/O (Gbps) Network (relative to non-private) $ relative to non- private Non-private 1x 1x CPIR 11.6 64 5x 265x Popcorn (delay 15s) 0.74 0.23 2x 3.87x ITPIR 3.1 64 2x 256x ITPIR++ (delay 15s) 0.65 3 2x 14x

slide-23
SLIDE 23

Popcorn is private and affordable but …

  • Assumes that the ITPIR servers do not collude.
  • Incurs costs that are linear in the size of the library.
  • Does not support recommendations, aggregate

view statistics.

Solution: Use prior work [Canny S&P ’02, Toubiana et al. NDSS

‘10]

slide-24
SLIDE 24

Related work

  • Improving performance of PIR.
  • Distributing work [FC13,

TDSC12], cheaper crypto [PETS16, ESORICS14, ISC10, TKDE13, WEWoRC07], bucketing [DBSec10, PETS10 ], batching [FC15, JoC04], secure co-processors [PET03, FAST13, NDSS08, IBM Systems Journal01]

  • Protecting library content in ITPIR [RANDOM98, S&P07, WPES13]
  • Handling variable-sized objects [CCSW14, NDSS13]
  • Prior PIR implementations [Percy++, PETS16, CCSW14]
  • Video-on-demand [MMCN95]
slide-25
SLIDE 25

Take-away points from Popcorn

  • It is possible to build a private, functional, and low-

cost media delivery system …

  • … by tailoring PIR to media delivery.
  • The per-request cost in Popcorn is 3.87x that of a

non-private baseline.