[Paper Notes] Some paper notes on the incentive system of blockchain

A Recommendation Scheme with Reputation-Based Incentive Mechanism on Consortium Blockchain

1. The main problems and solutions to be solved in the article:

1) Problems to be solved

Current recommendation engines are usually built on centralized servers. This architecture is vulnerable and exposes privacy. Therefore, a new decentralized architecture model is needed to solve the privacy protection problem of recommender systems.
A previous article used it
. However, the decentralized recommendation system faces several problems such as non-active participation and malicious behavior of nodes. These problems greatly affect the accuracy of recommendation results and the effectiveness of recommender systems. Therefore, this article will design an incentive mechanism to ensure the credibility and active participation of participants' recommendations.

2) The main work of the article

1. Propose a recommendation scheme based on the alliance blockchain
2. Introduce a reputation-based incentive mechanism
3. Use the Stackelberg game

The platform-centric mechanics of Stackelberg games assume that the platform has absolute control over the total amount paid by users, and that users can only increase their income by adjusting their strategies of participating in the system.

Our recommendation model is similar to the Stackelberg game model, but consists of a leader and multiple followers. In a recommendation scheme, the system is the leader and can formulate policies. Participants are followers, who can send recommendation requests as users, or choose as recommenders to make recommendations according to the leader's strategy.

The alliance blockchain recommendation scheme based on the reputation incentive mechanism uses the auditability and transparency of the alliance blockchain to enhance the credibility of the recommendation system.

2. System framework design

early stage
Multiple organizations can join forces to build a blockchain network.
After the consortium blockchain authentication and authorization, participants in the organization can join the blockchain network as nodes to perform recommended transactions.
start
The system platform publishes the tasks submitted by users to the blockchain network.
The reputation incentive mechanism attracts more nodes to participate in the recommendation.
When the number of participating nodes exceeds the minimum number specified by the system, the system will execute the recommendation.
Finish
Finally, the system returns the recommendation result to the user and distributes rewards to participating nodes. Malicious nodes will be punished.

3. Reputation Incentive Mechanism

Incentives

R is the total reward for all participants, α is the threshold of participating nodes required for the recommendation, and D is the deposit

Node reputation evaluation

Provide a deposit mechanism for new nodes to quickly increase credit. The more deposits, the higher the credit.

reward distribution

The recommendation scheme consists of two roles. One is the participants who recommend resources, pay a certain deposit Di, and get rewards after completing the recommendation; the other is users, who post recommendations and pay a certain reward R to all participants.

Definition of credible node:
if the reputation value T of the alliance node exceeds the credit threshold α, it is a credible node. The node is a malicious actor.

Design and implementation of recommender systems

We have united multiple educational institutions to form an educational alliance blockchain. The system is based on the Fabric development environment, and the operating system uses Ubuntu 64-bit system. The front-end system provides users with a visual interface for interacting with smart contracts, and communicates with background processes by calling RESTful APIs to realize system functions.

Users are divided into: ordinary users and joint node users.

The deposit part is generally the unreturned deposit, including the deposit for uncompleted recommendation tasks and the deposit deducted for malicious recommendation by nodes. The balance part is the bonus earned through referrals. The user will transfer different amounts of reward funds to the participant's balance account according to the different performance of the participating nodes.

4. Security and Performance Analysis

Anti-collusion analysis

System Robustness Analysis

With the advantages of blockchain, there will be no such risk.

Performance Testing

It can be clearly seen from the figure that after introducing incentives, the number of participants in each round of recommendation system increases significantly. We then consider ways to set thresholds to ensure a sufficient number of participants to improve the accuracy of recommendation results. Therefore, we conduct user satisfaction simulation analysis on the recommendation results under different thresholds. Let α be 50 70 90. As shown in Figure 6, when the number of participants increases, the user's satisfaction with the recommendation results also increases, and when the number of participants reaches a threshold, we can obtain more accurate recommendation results and thus higher user satisfaction. Therefore, when we design the incentive mechanism, we need to let the number of participants exceed the threshold.

Finally, the system incentive overheads of Optimal Recommendation Incentive Mechanism (ORIM) and Reverse Auction-based Dynamic Pricing Mechanism, Virtual Participation Credit (RADP-VPC) Incentive Mechanism and Reputation-Based Participatory Incentive Mechanism (RPIM) are compared.

The results of the system incentive overhead experiment in the case of collusion are shown in Figure 8. When there is collusion, the system incentive overhead of RADP-VPC increases rapidly, indicating that this mechanism cannot prevent malicious collusion. RPIM's system incentive fee increases slowly and steadily, indicating that this mechanism can guarantee the fairness of the system to a certain extent. The system incentive cost of ORIM is relatively stable, indicating that this mechanism ensures the fairness of the system and keeps the system incentive cost at a low level.

Incentive Mechanism for Reliable Federated Learning: A Joint Optimization Approach to Combining Reputation and Contract Theory

We design a reputation-based worker selection scheme for robust federated learning by using a multi-weight subjective logistic model. Leverage blockchain to achieve secure reputation management for employees with non-repudiation and tamper-proof properties in a decentralized manner. Furthermore, an effective incentive mechanism is proposed that combines reputation with contract theory to motivate high-reputation mobile devices with high-quality data to participate in model learning.

Why are incentives needed?

1. Unreliable workers may perform bad behaviors to mislead training. For intentional behavior, worker threads may launch poisoning attacks, sending malicious updates to affect global model parameters, causing current collaborative learning mechanisms to fail.
2. Self-serving mobile devices are reluctant to participate in model training without well-designed financial compensation.

system structure

Step 1 (Issue joint learning tasks and contract items)
Each task issuer broadcasts to mobile devices that it has specific resource requirements. Mobile devices meeting the requirements can become candidate models for training auxiliary roles to join federated learning tasks, and also feed back responses containing resource information to the task issuer.
Step 2 (Compute Candidate Reputation)
Step 3 (Select Workers for Federated Learning)
Step 4 (Perform Federated Learning and Evaluate Quality of Local Model Updates)

reputation design

In this section, we apply a subjective logistic model to generate synthetic reputation values ​​for worker candidates.

Record and Reward Federated Learning Contributions with Blockchain

existing problems

Although federated learning allows participants to contribute local data without revealing local data, it has issues with data security and accurate payment to participants for high-quality data contributions.

solved problem

A public blockchain with no transaction fees further incentivizes users to use it.

Incentive Mechanism for Reliable Federated Learning: A Joint Optimization Approach to Combining Reputation and Contract Theory

Direct reputation opinions generated from past interactions and indirect reputation opinions of other task issuers are integrated into a synthetic reputation for worker selection.

The response behavior of customers participating in the game is analyzed through the solution of the Stackelberg game and the effectiveness of our proposed framework is established through a case study.