Mehrdad Nojoumian Department of Computer & Electrical Engineering and Computer Science Florida Atlantic University 4 th AAAI Workshop on Incentives and Trust in E-Communities (WIT-EC) January 26, 2015 Trust, Influence and Reputation Management Based on Human Reasoning

Mehrdad Nojoumian Work-in-Progress  Interdisciplinary project between the Department of Computer and Electrical Engineering and Computer Science at Florida Atlantic University and the Department of Psychology at Southern Illinois University: Mehrdad Nojoumian, Assistant Professor Department of CEECS, Florida Atlantic University Dr. Meera Komarraju, Professor and Chair Department of Psychology, Southern Illinois University Colleen Bader, Graduate Student of Applied Psychology Department of Psychology, Southern Illinois University Kristiana Feeser, Graduate Student of Applied Psychology Department of Psychology, Southern Illinois University 2

Mehrdad Nojoumian Review of a Well-Known Trust Model  Previous Solution: trust value T(p+1) is given by the following equations and it depends on the previous trust rating where:   0 and   0 [CIA’00]. 3 T(p)CooperationDefection > 0 T(p) +  ( 1-T(p) )( T(p) +  ) / ( 1- min{ |T(p)|, |  | } ) < 0 ( T(p) +  ) / ( 1 - min{ |T(p)|, |  | } ) T(p) +  ( 1+T(p) ) = 0  T(p) < 0 T(p) = 0 T(p) > 0 T(p) < 0 T(p) = 0 T(p) > 0

Mehrdad Nojoumian Our Sample Trust Model  Our Function is not just a function of a single round, but of the history: Reward more (or same) the better a participant is, Penalize more (or same) the worse a participant is. 4 DiscourageReward T Good P i  ( ,+1] OpportunitiesGive/Take T New P i : [ ,  ] PenalizeEncourage T Bad P i  [-1,  ) DefectionCooperation Trust Value Bad P i (C) Newcomer P i (C) Good P i (C) Bad P i (D) Newcomer P i (D) Good P i (D)

Mehrdad Nojoumian Intuition Behind Our Model 5 There exist some common principles for trust modeling AB A lies to B for the 1 st time: defection AB A lies to B for the 2 nd time: same defection + past history AB A cheats on B: costly defection Impact of the brain’s history

Mehrdad Nojoumian Contents of the Talk 6  Our Motivation and Objectives  Trust, Influence and Reputation  Data Collection Methodology  Sample Scenarios and Framing  Templates of Our Scenarios  Trust Modeling in Four Steps Specification, Transformation, Evaluation and Modification

Mehrdad Nojoumian Our Motivation and Objectives  Trust, influence & reputation are human factors that have been widely used in technological systems. We intend to scrutinize them from both angles.  Our intention is to construct new computational models of these notions while incorporating human reasoning factors into the models’ specifications.  We intend to utilize our models in different contexts such as cryptographic constructions and cybersecurity where players are humans, e.g., hackers.  Objectives: 1.Bridge the gap between computer and social sciences. 2.Investigate how trust measurement and influence work in humans. 3.Perform cross-cultural data collection targeting East and West. 4.Quantify our data through hypothesizing and modeling. 5.Finally, deploy our models in technological and software systems. 7

Mehrdad Nojoumian Trust, Influence and Reputation  Social Science Perspective: Trust is the willingness of a person to become vulnerable to the actions of another person irrespective of the ability to control those actions. Influence refers to any tactic used to alter the behavior, decision or attitude of other people.  Computer Science Perspective: Trust is defined as a personal expectation that a player has with respect to the future behavior of another party, i.e., a personal quantity. Reputation is the perception that players have with respect to another player’s intention, i.e., a social quantity. 8

Mehrdad Nojoumian Data Collection Methodology  First, demographic data will be collected at the beginning of the data collection to obtain a profile of our human subjects. Then, questions are asked from our subjects based on 8 distinct groups: 1.Initial Trust: scenarios that illustrate the initial trust values in the first dyadic encounters, e.g., when two people meet for the first time or after online chatting, etc. 2.Trust Escalation: scenarios that illustrate a sequence of incidents/actions that escalate trust between two parties, e.g., helping, supporting, lending money, etc. 3.Trust Reduction: scenarios that illustrate a sequence of incidents/actions that damage trust between two parties, e.g., lying, lying for the second time, cheating, etc. 4.Trust Mutation: 2 nd, 3 rd groups can be a sequence of mild incidents (lying) followed by critical incidents (cheating) and vice versa. 9

Mehrdad Nojoumian Data Collection Methodology (Cont.) 5.Re-Building Trust: scenarios that demonstrate how trust can be re- built between two parties, e.g., lying, ignoring and then apologizing, paying attention, supporting, etc. 6.Gaining Influence: scenarios that illustrate how people might be able to change the attitude of others and the ratio of this impact over time, e.g., helping, lending money, providing useful advice, etc. 7.Losing Influence: scenarios that illustrate how people may lose their influential impacts on others and the ratio of this failure over time, e.g., forcing, misleading, etc. 8.Influence Mutation: 6 th, 7 th groups can be a sequence of mild incidents (misleading) followed by critical incidents (forcing) and vice versa.  We have defined a list of positive and/or negative keywords: +lending, +helping, +supporting, +giving, -lying, -cheating, -ignoring, -misleading, +/- judging, +/- relying, +/- suggesting, etc. 10

Mehrdad Nojoumian Sample Scenario  Measurements: Untrustworthy (U) High Negative (H-) Medium Negative (M-) Low Negative (L-) Neutral (N) Low Positive (L+) Medium Positive (M+) High Positive (H+) Trustworthy (T) 11

Mehrdad Nojoumian Framing in Scenarios  Actual Scenarios: Hanna and Mary have known each other for about five years and are close friends. They were roommates during college and graduated with a Bachelor’s degree in business from UCLA in Los Angeles, California. They were delighted the company that hired them placed them with the same supervisor. As they began their employment, they decided to be roommates once again. 1.If you were Mary, how much trust would you have in Hanna after graduating from college? (Initial Trust) Complete Distrust Some Distrust Neither Distrust/Trust Some Trust Complete Trust After working together for about a year, Mary started noticing some changes in Hanna’s behavior. For instance, once, when Mary was ill and missed work on a day when there was a staff meeting, Hanna told her that she had not missed anything important. Later, Mary found out that Hanna had failed to inform her about some changes regarding a deadline for Project 1. Finally, one morning, as Mary prepared to leave for work, she realized she could not find the presentation that she and Hanna had carefully prepared the night before. She also noticed that although she and Hanna always car pooled to work, it looked like Hanna had already left for work without her. 2.If you were Mary, how much trust would you have in Hanna after you found out that she had left for work without you? (Trust Reduction) U H − M − L − N L + M + H + T 12

Mehrdad Nojoumian Templates of Our Scenarios 1.Stranger or Weak relationship, e.g., business scenarios. 2.Friends or Medium relationship, e.g., colleague scenarios. 3.Partners or Close relationship, e.g., romantic scenarios. 13 Templat-1Templat-2Template-3Template-4Templat-5Templat-6Templat-7Templat-8 Initial Trust Trust Reduction Trust Escalation Trust Reduction Further Trust Reduction Further Trust Escalation Trust Reduction Trust Escalation Trust Mutation (negative) Trust Mutation (positive) Trust Mutation (negative) Trust Mutation (positive) Trust Mutation (negative) Trust Mutation (positive) Trust Mutation (negative) Rebuild Trust Strengthen Trust Rebuild Trust

Mehrdad Nojoumian Trust Modeling in Four Steps 1.Specification: If a bad party cooperates, he is encouraged by a small reward. If a newcomer cooperates, he is rewarded. If a good party cooperates, he is rewarded by a factor more than the encouragement factor. If a good party defects, he is discouraged by a small penalty. If a newcomer defects, he is penalized. If a bad party defects, he is penalized by a factor more than the discouragement factor. 1.Transformation: 14

Mehrdad Nojoumian Trust Modeling in Four Steps (Cont.) 3.Evaluation: Behavioral: how the model performs among a large enough number of players by running standard tests, i.e., a sequence of “C” and “D” for each player. Adversarial: how vulnerable the model is to different attacks, e.g., the “Sybil attack” where the system is subverted by forging identities. 4.Modification: If all parties have cooperated, it is not required to increase the trust values or if all players have defected, it is not required to decrease the trust values. If majority of the players have cooperated, cooperation should be rewarded less and defection should be penalized more. If majority of the players have defected, defection should be penalized less and cooperation should be rewarded more. If the number of cooperative and non-cooperative players are equal, cooperation and defection should be readjusted with an equal ratio. 15

Mehrdad Nojoumian Thank You Very Much Questions? 16