Paper Information

DOI: 10.1109 / TCOMM.2020.2990686

1. Summary

The author proposes Blockchain-based Federated Learning (BFL) for privacy awareness and efficient vehicle communication networks. Among them, the on-board machine learning (oVML) model of autonomous vehicles is updated for distributed exchange and verification. (Data transaction with neighboring vehicles)
bonus system.
A mathematical framework was developed: characterized by controllable network and BFL parameters (such as: retransmission limit, block size, block generation rate, frame size) in order to capture their impact on system-level performance.
The end-to-end delay is quantified.

2. Background

The new generation of wireless networks is more reliable and will soon meet the real-time communication needs of autonomous vehicles. There are many self-driving cars, and traditional cloud methods are riskier (high-latency, cloud-based learning methods are slow, and the delay for completion of GFL training is about 10 minutes), while autonomous driving requires low latency.

Data transactions between vehicles should ensure the privacy of the original data to other adjacent vehicles.

Limitations of traditional FL:

(1) Too much dependence on the central server.

(2) Equipment credibility. Without an incentive mechanism, local equipment does not actively participate in training.

3. This article contribution

(1) Propose BFL and replace the central server with blockchain.

(2) Do a mathematical analysis of the end-to-end delay.

(3) A similar vehicle-mounted network model was constructed, which combined the complex interaction of wireless channel and link characteristics with the detailed transmission mechanism of the cellular network, and successfully measured the impact on BFL. A lot of numerical and simulation experiments have been done to evaluate the proposed framework under different network scenarios and channel conditions.

(4) Due to the real-time changes in the network settings of autonomous vehicles (the number of vehicles, miners, wireless channel conditions, etc.), the author designed an online algorithm to minimize the system delay under real network conditions. The algorithm continuously monitors the deviation of the system delay from the target value, and adjusts the block generation rate to ensure the minimum total delay, and can effectively cope with the dynamic network.

4. BFL model details

4.1 Model overview

The picture above is an abstract diagram of the BFL framework composed of self-driving cars and miners.

Miners are physically movable vehicles or independent nodes at the edge of the network (such as cellular base stations). Vehicles and miners use a uniform random method to trust each other and communicate anonymously. The benefits are:

(1) Reduced malicious attacks in fixed connections.

(2) Incentive mechanism. Don't make assumptions about their honesty.

(3) It is easier to make the vehicles in the BFL form a representative group.

(4) it allows the global model and findings to apply to the entire BFL System. (This sentence is not very understanding)

The interaction in the figure above is explained as follows.

Each oVML performs local learning and sends local model updates to the miners associated with it in the network; all miners exchange and verify all their local model updates, and then perform PoW. When the miner completes his PoW, he generates a block by recording the verified local model update. Finally, the block is inserted into the distributed ledger of the blockchain. Then, all relevant vehicles can use this newly inserted block to calculate the required global model update.

The network system can exchange local models of vehicles, while providing verification and corresponding rewards.

The global model update for each vehicle is calculated locally, but will cause additional delays (due to the blockchain).

4.2 Model problems and solutions

This article solves a parallel regression problem.

$S = \bigcup_{n=1}^{N}S_{n}$ All devices in the network.

Training objectives: for a global vector $v$ , $Minimize F\left ( v \right )$ which $F(v) = \frac{\sum_{n=1}^{N}\sum_{d_{j}\epsilon s_{n}}(a_{j}^{T}v-b_{j})^{2}/2}{\left | S \right |}$

$Minimize F\left ( v \right )$ The idea of the solution is to perform local training at each oVML through the stochastic gradient algorithm, followed by global training through the distributed Newton method to gather updates. At each stage, the oVML local model is updated as follows after each iteration:

among them,

[Some explanations]

1. Each module in the ledger consists of two parts: header and body.

The header contains 3 parameters: 1) pointer to the previous block, 2) block generation rate, 3) PoW result.

The body contains: 1) local calculation delay, 2) local update $\left ( v_{n} , \bigtriangledown F_{j}\left ( v \right )\right )$ .

2. In this article, the main factors for forking: 1) block generation rate, 2) fast propagation delay.

3. Reward mechanism.

(1) Data rewards. Provide data sample rewards to vehicles. (The reward is proportional to the size of the local data sample)

(2) Mining rewards. Provide rewards for the block verification process to the victory miners. (The reward is proportional to the mining power contributed by the miner, and the mining power of the miner is determined by the set of useful data samples associated with the oVML)

[Note] 1) PoW is used for proof of work. The author said that the strong relationship between Byzantine Fault Tolerance and BFL proved to be a bit too "too simple", which may require more complicated calculations and prologues to reach agreement among miners.

2) General Motors, BMW, etc. have begun to use blockchain as a way to share self-driving car data between them and other manufacturers.

5. Delay analysis

Calculate proof of work, communication and calculation delays. Analyze the end-to-end total delay and dynamically adjust the block generation rate to minimize the overall perceived delay of the system.

XX. Conclusion

BFL is effective.
The end-to-end delay leads to the optimal block generation rate.
An optimization method based on usage analysis strikes a good balance between omitting specific model details and considering key system features.

XX. Potential future research directions

(1) Considering the automatic characteristics of autonomous vehicles used for distributed learning, develop complex mobility models.

(2) Verify the risk analysis and solution design of privacy leakage during training.

(3) Research on a blockchain system based on context-aware integrated (and robust) distributed learning for ultra-reliable low-latency applications.

[To be continued..................]

[Notes] "Federated Learning With Blockchain for Autonomous Vehicles Analysis and Design Challen" intensive reading notes