1. Private Matchings and Allocations
Joint work with
Justin Hsu (Penn)
Zhiyi Huang (HKU)
Aaron Roth (Penn)
Tim Roughgarden (Stanford)
Speaker: Steven Wu
University of Pennsylvania
STOC 2014
6. Differential Privacy
• An algorithm A with domain X and range R
satisfies ε-differential privacy if for every outcome r
and every pair of databases D, D’ differing in one
record:
Pr[ A(D) = r ] ≤ (1 + ε)Pr[ A(D’) = r]
• Domain: Reported valuation functions
• Range: Matchings
11. Supply Assumption
• We need multiple copies for each
type of good even under JDP.
• How many?
Impossible Trivial
12. Main Result
Theorem: There is a JDP algorithm in the that solves the
max-weight matching problem with n people and k
types of goods with supply at least s each, and
outputs a matching of weight
OPT – αn
whenever:
14. “Low information” Signal
o From the signal, every bidder can figure out what
item they are matched to in a matching
o Does not reveal each individual’s private data
• Think: Prices
15. Max Matchings (A Sketch)
A remarkable algorithm for Max-Matchings: [Kelso and Crawford ’82]
18. Prices as information
Claim: Bidders just need to see the prices
1. Prices are sufficient to identify the favorite good
2. When price raises again, a bidder is unmatched
3. Bidders are matched to the last thing they bid on
• Just need to count how many bids each good
received!
19. Privately Maintaining Counts
1 0 0 1 1 1 0 1 032
• Private (noisy) counters under continual observation
[DNPR10, CSS10]
• Given a stream of T bits, maintain an estimate of the
running count with accuracy
o Single Stream of sensitivity 1
20. Privately Maintaining Counts
1 1 1 1 1 1 0 0 018
1 0 0 1 1 1 0 1 032
0 0 0 1 1 1 1 1 0192
• A straightforward generalization:
K counters on K streams that collectively have
sensitivity Δ gives accuracy
23. Supply
• Goods might also be under/over-allocated by E.
o Doesn’t reduce the welfare by more than (1-α) factor if
24. Main Theorem
Theorem: There is a private algorithm in the billboard
model that solves the max-weight matching problem
with n people and k types of goods with supply at
least s each, and outputs a matching of weight
OPT – αn
whenever:
26. Conclusions
• Some problems that can’t be solved under DP
can be solved under joint-DP.
o If the output is partitioned among the agents
o The agent’s output is allowed to be sensitive in his input.
• Billboard model: interesting framework to design
a joint-DP algorithm?
27. Private Matchings and Allocations
Joint work with
Justin Hsu (Penn)
Zhiyi Huang (HKU)
Aaron Roth (Penn)
Tim Roughgarden (Stanford)
Speaker: Steven Wu
University of Pennsylvania
Hinweis der Redaktion
Of course, since our talk is the last talk of this session. The first notion of privacy we would look at is differential privacy. Some of you might have seen this pictures many times. It says that if we change the input database by one record, the output distribution induced by our randomized algorithm is not affected by much.
If there is no contention for goods, a high welfare matching has to give people what they want. A high welfare allocation will give people what they want. Consider a scenario where people either like alcohol or cigarettes. The tension between social welfare and privacy makes it impossible to work with standard differential privacy.
Not the songs you heard on radio. A nice decentralized model
The idea is to have our mechanism broadcast some low information signal to all of the agents. The signal itself should satisfy standard DP. And a natural candidate for such a signal in allocation problem is prices.