Related concepts of federal learning and motivation (1)

Paper title ：Incentive Mechanisms for Federated Learning: From Economic and Game Theoretic Perspective

The incentive mechanism design of federal learning ： Concept definition and motivation

stay FL In the scene , Participants may be reluctant to participate in training without compensation, because it will lose resources to train models in vain and bear the risk of privacy disclosure . meanwhile , Incentive mechanism can also reduce information asymmetry （server and worker） The negative impact . An excellent incentive mechanism may have the following characteristics ：

Incentives can be coordinated 、 trusted ： Every worker Can get the best compensation , As long as they work honestly ; in other words , If they do evil, they will not increase their profits

Personal rationality ： in other words worker Participate in FL The return is nonnegative
Bill balance ： Yes workers The total payment will not be greater than the given budget
The calculation is valid ： In polynomial time , The incentive mechanism can complete worker Election and distribution of awards
Fairness ： When the predefined fairness equation （ Contribution equity ） When it reaches its peak , Incentive mechanism can achieve fairness . A fair incentive mechanism can optimally distribute rewards

About FL Some definitions of incentive mechanism in ：

（p,c,r）
p： Participant , They provide useful training resources
c： Used to measure each worker One way to contribute
r： be based on c Of , For each worker The way to give rewards
Specially , The purpose of designing incentive mechanism is worker The optimal level of participation and The best reward To sustain FL The sustainability of
So , The key to the incentive mechanism is Contribution assessment and Bonus distribution

Assessment of contributions

stay FL in , If you can get higher rewards , self-interested DO There will be a higher willingness to join FL; However , From another perspective , It's right MO Cause greater financial consumption . therefore , Design contribution evaluation is needed to balance . The literature 22 Shows about honesty DO The contribution of 、 malice DO Analysis of behavior and attack oriented defense mechanism ; The literature 23 Adopted Attention mechanism To evaluate longitudinal FL Medium DO The gradient contribution of . This method , For each DO Measure real-time contributions , Have high sensitivity to the quantity and quality of data . The literature 24 An intuitive contribution evaluation method based on step-by-step contribution calculation is proposed . The literature 25 in , The author proposes a contribution evaluation method based on reinforcement learning . Specially , The literature 26 A new method called “ be based on peer Predicted pairwise correlation protocols ” stay No test set In the case of FL User contributions in , It uses user uploaded ** About the statistical correlation of model parameters “ To make a specific assessment .

However ,22-24 All methods assume a premise , That is, there is a credible Center server Will honestly calculate each DO The contribution of , This assumption will lack transparency and then hinder the actual FL The success of the . To solve this problem , Blockchain based p2p Payment system （27-28） It is proposed to support the adoption of consensus agreements based on SV To replace the traditional third party . meanwhile , In order to prevent malicious behavior ,29 The author proposed a framework based scoring rule to promote DO Upload their models reliably .

at present FL in Contribution assessment The mainstream strategies of can be divided into the following ：

• Self reporting based on contribution assessment ： This is the most straightforward way , This is DO Take the initiative to MO Report on your contributions . In this scenario , There are many advantages , For example, the scale of computing resources and data scale （ I think it means DO It will be much more convenient to count by yourself , But it still has the possibility of false report ）
• Based on contribution evaluation Shapley Value： This is a method of considering edges , It will DO The influence of the order of joining is taken into account , So as to fairly count their marginal effects . This method is usually used with ”cooperatetive game” sv Is defined as follows ：
  This expression means removing i All of the DO The average edge contribution inside ,S It represents the alliance N Different cooperation modes in ,v（s） It's a subset s The utility of the jointly trained model , lately 33-36 The improvement of this edge model is described
• Based on contribution evaluation influence and reputation： One worker Of influence Defined as its right FL The contribution of the loss function of the model . Through the update of model or data , The loss function will be improved . The literature 38 Put forward a new concept ,Fed-Influnce, It is mainly used to quantify each individual client Of , Not model parameters , At the same time, it can perform well on convex and non convex functions .reputation The mechanism is mainly combined with blockchain to elect reliable worker（39-42）.DO Of reputation It can be divided into direct reputation And recommended reputation, Then use subjective logic model to calculate .

Distribution of awards

After the evaluation DO After your contribution ,MO It should be right DO Distribute incentives to retain and improve the availability of high-quality data DO The number of

• Offered Reward ： This method considers MO stay DO Reward before you finish training , The reward can be based on the quality of the resources provided （44）, Or by voting （45） To decide
• payoff sharing： This method considers Mo stay Do After completing the task, based on the reward . But what? , Such delayed payment will reduce worker Enthusiasm ,19-20 The literature suggests payoff sharing Can dynamically allocate established budget. The goal of this method is to solve a value Reduced regret Optimization of mobile , Contribution fairness can be achieved 、regert Distribution fairness , Expect fairness, etc .

summary

In this session , We are right. FL Training process of 、 Basic framework 、 Introduce the advantages . Besides ,FL The basis of incentive mechanism is also discussed . for example , Concept definition and motivation . Next time we will show some basic economic and game theory models .