Mechanism Design for Mobile Phone Sensing Dejun Yang, Guoliang - PowerPoint PPT Presentation

Crowdsourcing to Smartphones: Incentive Mechanism Design for Mobile Phone Sensing Dejun Yang, Guoliang (Larry) Xue, Xi Fang and Jian Tang Arizona State University Syracuse University

Global Smartphone Users >2B 1.08B 500M 2010 2012 2015 2/36 Date Source: IDC http://www.idc.com/getdoc.jsp?containerId=233553, Go-Gulf http://www.go-gulf.com/blog/smartphone Image source: http://www.foxshop.seeon.com/images/smartphone_shadow-group.jpg

Mobile Phone Sensing Apps 3/36 Image source: http://www.mynewplace.com/blog/files/2011/05/smart-phone-user.jpg, http://serc.carleton.edu/images/sp/library/google_earth/google_maps_new_york.v2.jpg, http://media.treehugger.com/assets/images/2011/10/nextfest-peir-001.jpg

What is Missing? Smartphone users consume their own resource CPU Power Memory 4/36

Related Works • Developed recruitment frameworks • Focused on user selection, not incentive design S. Reddy D. Estrin M.B. Srivastava • Developed a sealed-bid second-price auction • The platform utility was not considered G. Danezis S. Lewis R. Anderson • Designed an auction based dynamic price incentive mechanism • Truthfulness was not considered J-S. Lee B. Hoh S. Reddy, D. Estrin, and M.B. Srivastava ; “Recruitment framework for participatory sensing data collections” in PERVASIVE 2010 G. Danezis , S. Lewis, and R. Anderson; “How Much is Location Privacy Worth ?” In WEIS 2005. 5/36 J-S. Lee and B. Hoh; “Sell Your Experiences: Market Mechanism based Participation Incentive for Participatory Sensing” in PERCOM 2010

Other Related Works MAUI Energy Saving (MobiSys 2010) PRISM Application (MobiSys 2010) Development PEPSI Privacy (WiSec 2011) TP (HotNets 2011) 6/36

Outline/Progress Related Works System Model Platform-Centric Model User-Centric Model Simulation Results Conclusions 7/36

System Model 𝑉 = {1, 2, … , 𝑜} , 𝑜 ≥ 2 Platform-Centric Model User-Centric Model 8/36

Platform-Centric Model • Platform announces a total reward 𝑆 • Each user 𝑗 has the sensing time 𝑢 𝑗 ≥ 0 and sensing cost 𝜆 𝑗 × 𝑢 𝑗 , where 𝜆 𝑗 is its unit cost • The utility of user 𝑗 is 𝑢 𝑗 𝑣 𝑗 = 𝑆 − 𝑢 𝑗 𝜆 𝑗 𝑢 𝑘 𝑘∈𝑉 • The utility of the platform is 𝑣 0 = 𝜇 log 1 + log 1 + 𝑢 𝑗 − 𝑆 𝑗∈𝑉 where 𝜇 > 1 is a system parameter. 9/36

User-Centric Model • Platform announces a set Γ = {𝜐 1 , 𝜐 2 , … , 𝜐 𝑛 } of tasks, where each 𝜐 𝑘 has a (private) value 𝜉 𝑘 > 0 . • Each user 𝑗 ∈ 𝑉 selects a subset Γ 𝑗 ⊆ Γ , based on which user 𝑗 has a (private) cost 𝑑 𝑗 𝑇 Auction Γ 1 , 𝑐 1 , … , Γ 𝑜 , 𝑐 𝑜 𝑞 1 , 𝑞 2 , … , 𝑞 𝑜 𝑗 = 𝑞 𝑗 − 𝑑 𝑗 , 𝑗𝑔 𝑗 ∈ 𝑇, • Utility of user 𝑗 is 𝑣 0, 𝑝𝑢ℎ𝑓𝑠𝑥𝑗𝑡𝑓. 0 = 𝜉 𝑇 − • Utility of the platform is 𝑣 𝑞 𝑗 ,where 𝑗∈𝑇 𝜉 𝑇 = 𝜉 . 𝑘 𝜐 𝑘 ∈∪ 𝑗∈𝑇 Γ 𝑗 10/36

Mobile Phone Sensing System Sensing Task Desc. Sensing Plan Sensed Data Platform Smartphone Users 11/36

Outline/Progress Related Works System Model Platform-Centric Model User-Centric Model Simulation Results Conclusions 12/36

Stackelberg Game (Platform-Centric) Leader Followers Stackelberg Equilibrium :  Each follower tries to maximize its utility, given the leader’s strategy  The leader tries to maximize its utility, given the knowledge of the followers’ behavior 13/36

User Sensing Time Determination Sensing Time Determination (STD) game: Leader Players: Users Strategy: Sensing Time Followers 𝑢 𝑗 Utility: 𝑣 𝑗 = 𝑆 − 𝑢 𝑗 𝜆 𝑗 𝑢 𝑘 𝑘∈𝑉 14/36

NE Computation Sort users according to their unit costs, 𝜆 1 ≤ 𝜆 2 ≤ ⋯ ≤ 𝜆 𝑜 . 𝑇 ← {1, 2} , 𝑗 ← 3 ; 𝜆 𝑗 + 𝜆 𝑘 𝑘∈𝑇 while 𝑗 ≤ 𝑜 𝑏𝑜𝑒 𝜆 𝑗 < Leader |𝑇| 𝑇 ← 𝑇 ∪ 𝑗 , 𝑗 ← 𝑗 + 1; end for each 𝑗 ∈ 𝑉 Followers 𝑜𝑓 = 𝑇 −1 𝑆 𝑇 −1 𝜆 𝑗 if 𝑗 ∈ 𝑇 then 𝑢 𝑗 1 − ; 𝜆 𝑘 𝜆 𝑘 𝑘∈𝑇 𝑘∈𝑇 𝑜𝑓 = 0 ; else 𝑢 𝑗 𝑜𝑓 𝑜𝑓 , 𝑢 2 𝑜𝑓 , … , 𝑢 𝑜 return 𝑢 1 T HEOREMs 1&2: The strategy profile 𝑢 𝑜𝑓 = 𝑢 1 𝑜𝑓 is 𝑜𝑓 , 𝑢 2 𝑜𝑓 , … , 𝑢 𝑜 the unique NE of the STD game. 15/36

Platform Reward Determination 𝑣 0 = 𝜇 log 1 + log 1 + 𝑢 𝑗 − 𝑆 𝑗∈𝑉 Leader 𝑣 0 = 𝜇 log 1 + log 1 + 𝑌 𝑗 𝑆 − 𝑆 𝑗∈𝑇 𝑇 −1 𝑇 −1 𝜆 𝑗 Followers where 𝑌 𝑗 = 1 − 𝜆 𝑘 𝜆 𝑘 𝑘∈𝑇 𝑘∈𝑇 T HEOREM 3: There exists a unique SE R ∗ , 𝑢 𝑜𝑓 in the MSensing game, where 𝑆 ∗ is the unique maximizer of the above utility function, which is strictly concave. 16/36

Outline/Progress Related Works System Model Platform-Centric Model User-Centric Model Simulation Results Conclusions 17/36

LSB Auction (Not Truthful) 𝑇 ← 𝑗 , where 𝑗 ← arg max i∈𝑉 𝑔 𝑗 ; 𝜗 ⋇ while ∃𝑗 ∈ 𝑉 ∖ 𝑇 such that 𝑔 𝑇 ∪ 𝑗 > 1 + 𝑜 2 𝑔 𝑇 𝑇 ← 𝑇 ∪ {𝑗} ; 𝜗 if ∃𝑗 ∈ 𝑇 such that 𝑔 𝑇 ∖ 𝑗 > 1 + 𝑜 2 𝑔 𝑇 𝑇 ← 𝑇 ∖ {𝑗} ; go to ⋇ ; if 𝑔 𝑉 ∖ 𝑇 > 𝑔 𝑇 then 𝑇 ← 𝑉 ∖ 𝑇 ; for each 𝑗 ∈ 𝑉 if 𝑗 ∈ 𝑇 then 𝑞 𝑗 ← 𝑐 𝑗 ; else 𝑞 𝑗 ← 0 ; return (𝑇, 𝑞) 0 𝑇 + 𝑔 𝑇 = 𝑣 𝑐 𝑗 is 𝑡𝑣𝑐𝑛𝑝𝑒𝑣𝑚𝑏𝑠 and nonnegative 𝑗∈𝑉 18/36

Truthful Auction T HEOREM 5: An auction mechanism is truthful if and only if, for any bidder i and any fixed choice of bid b -i by other bidders, 1)The selection rule is monotonically nondecreasing in 𝑐 𝑗 ; 2)The payment p i for any winning bidder i is set to the critical value. 19/36

MSensing Auction Winner Determination 𝑇 ← ∅ , 𝑗 ← arg max 𝑘∈𝑉 𝑤 𝑘 𝑇 − 𝑐 𝑘 ; while 𝑐 𝑗 < 𝑤 𝑗 and 𝑇 ≠ 𝑉 𝑇 ← 𝑇 ∪ {𝑗} ; Pricing 𝑗 ← arg max 𝑘∈𝑉∖𝑇 𝑤 𝑘 𝑇 − 𝑐 𝑘 ; 20/36

MSensing Auction Winner 𝑞 𝑗 ← 0 for all 𝑗 ∈ 𝑉 ; Determination for each 𝑗 ∈ 𝑇 𝑉 ′ ← 𝑉 ∖ {𝑗} , 𝑈 ← ∅ ; repeat 𝑗 𝑘 ← arg max j∈𝑉 ′ ∖𝑈 (𝑤 𝑘 𝑈 − 𝑐 𝑘 ) ; 𝑞 𝑗 ← max 𝑞 𝑗 , min 𝑤 𝑗 𝑈 − 𝑤 𝑗 𝑘 𝑈 − 𝑐 𝑗 𝑘 , 𝑤 𝑗 𝑈 ; Pricing 𝑈 ← 𝑈 ∪ {𝑗 𝑘 } ; until 𝑐 𝑗 𝑘 ≥ 𝑤 𝑗 𝑘 or 𝑈 = 𝑉′ ; if 𝑐 𝑗 𝑘 < 𝑤 𝑗 𝑘 then 𝑞 𝑗 ← max 𝑞 𝑗 , 𝑤 𝑗 𝑈 ; 21/36

Walk-through Example (MSensing) 8 6 6 5 4 1 2 3 1 2 3 4 5 6 3 8 6 8 10 9 Winner Selection : 𝑇 = ∅ : 𝑤 1 ∅ − 𝑐 1 = 𝑤 ∅ ∪ 1 − 𝑤 ∅ − 𝑐 1 = 19 , 𝑤 2 ∅ − 𝑐 2 = 18 , 𝑤 3 ∅ − 𝑐 2 = 17 𝑤 4 ∅ − 𝑐 4 = 1 . 𝑇 = {1} : 𝑤 2 1 − 𝑐 2 = 𝑤 1 ∪ {2} − 𝑤 1 − 𝑐 2 = 2 , 𝑤 3 1 − 𝑐 3 = 3 , 𝑤 4 1 − 𝑐 4 = −5 . 𝑇 = {1,3} : 𝑤 2 1,3 − 𝑐 2 = 𝑤 1,3 ∪ {2} − 𝑤 1,3 − 𝑐 2 = 2 , 𝑤 4 1 − 𝑐 4 = −5 . 𝑇 = {1,3,2} : 𝑤 4 1,3,2 − 𝑐 4 = −5 . 22/36

Walk-through Example (MSensing) 8 6 6 5 4 1 2 3 1 2 3 4 5 6 3 8 6 8 10 9 Payment Determination : 𝑞 1 : Winners are {2,3}. 𝑤 1 ∅ − 𝑤 2 ∅ − 𝑐 2 = 9 , 𝑤 1 {2} − 𝑤 3 2 − 𝑐 3 = 0 , 𝑤 1 2,3 = 3 . 𝑞 1 = 9 ≥8. 𝑞 2 : Winners are {1,3}. 𝑤 2 ∅ − 𝑤 1 ∅ − 𝑐 1 = 5 , 𝑤 2 {1} − 𝑤 3 1 − 𝑐 3 = 5 , 𝑤 2 1,3 = 8 . 𝑞 2 = 8 ≥6. 𝑞 3 : Winners are {1,2}. 𝑤 3 ∅ − 𝑤 1 ∅ − 𝑐 1 = 4 , 𝑤 3 {1} − 𝑤 2 1 − 𝑐 2 = 7 , 𝑤 3 1,2 = 9 . 𝑞 3 = 9 ≥6 . 23/36

MSensing is Truthful T HEOREM 6. MSensing is computationally efficient, individually rational, profitable and truthful. 24/36

Outline/Progress Related Works System Model Platform-Centric Model User-Centric Model Simulation Results Conclusions 25/36

Simulation Setup • Platform-Centric Model – 𝑜 is varied from 100 to 1000 – Cost is uniformly distributed over [1, 𝜆 𝑛𝑏𝑦 ] , where 𝜆 𝑛𝑏𝑦 is varied from 1 to 10 – 𝜇 is set to 3, 5, 10 • User-Centric Model – 𝑜 is varied from 1000 to 10000 – 𝑛 is varied from 100 to 500 – 𝜗 is set to 0.01 26/36

Platform-Centric Incentive Mechanism Running Time 27/36

Platform-Centric Incentive Mechanism Number of Participating Users 28/36

Platform-Centric Incentive Mechanism Platform Utility 29/36

Platform-Centric Incentive Mechanism User Utility 30/36

Simulation Setup • User-Centric Model 31/36

User-Centric Incentive Mechanism Running Time 32/36

User-Centric Incentive Mechanism Platform Utility 33/36

User-Centric Incentive Mechanism Verification of Truthfulness 𝑑 851 = 18 𝑑 333 = 3 34/36

Outline Related Works System Model Platform-Centric Model User-Centric Model Simulation Results Conclusions 35/36