Short video content classification method based on quantum federal learning

Abstract

The invention discloses a short video content classification method based on quantum federal learning, and the method is characterized in that the method comprises the following steps: a quantum server initializes a model, and independently measures each parameterized quantum state through a random Purlei matrix for multiple times, so as to construct a classical shadow chromatography; a classic edge client obtains interestingness vectors of the client for various types of short video contents based on a local model and unlabeled image data newly added into a photo album, and encodes the vectors into observable measurement. And then calculating a local gradient parameter according to the global gradient, the own data coding observed quantity and a loss function. And the quantum server aggregates and updates the global parameters. And iterating the above steps until a preset number of times is reached, terminating model training, and obtaining an interest vector of a final user group. The method has the beneficial effects that the advantages of quantum computing are utilized under the condition that edge equipment is not required to have quantum capability, the model training performance is improved, and meanwhile, the local data privacy of a user is protected.

Description

Technical Field

[0001] The present invention relates to the technical field of quantum federated learning, and in particular to a short video content classification method based on quantum federated learning. Background Art

[0002] Short videos are the main source of Internet traffic on the user side. Short video classification strategies that meet user interests often bring greater economic benefits to content providers. The short video content classification method based on federated learning technology is a distributed machine learning technology. Its core idea is to conduct distributed model training through multiple edge computing devices holding local data, and finally aggregate global parameters through a server with fusion computing capabilities. In this process, the edge computing device does not need to provide local data to the server, but only needs to exchange model parameters with the server, thereby realizing the privacy protection function. Quantum computing is an emerging computing method with great potential. The basic assumptions of quantum mechanics ensure that its computing power far exceeds that of classical computing methods. Servers with quantum computing capabilities have huge computing power advantages over classical servers. For edge clients that only have classical computing capabilities, the use of quantum state shadow tomography technology can eliminate the requirements of quantum computing frameworks for their quantum computing capabilities.

[0003] On the one hand, the short video content classification task requires the use of quantum computing capabilities to improve learning efficiency. On the other hand, existing edge devices do not have quantum computing capabilities and cannot directly adopt the full-process quantum computing model. Summary of the invention

[0004] To this end, the present invention provides a short video content classification method based on quantum federated learning to solve the problems of low efficiency and limited device capabilities. The design of the "quantum server-classical client" architecture can take advantage of quantum computing without requiring edge devices to have quantum computing capabilities.

[0005] In order to achieve the above object, the present invention provides the following technical solution: a short video content classification method based on quantum federated learning, comprising:

[0006] S1: The quantum server initializes the model with random model parameters θ0 and performs parameterized quantum state

[0007] Construct a classic shadow tomography SG0, where

[0008] Represents θ k is θ k ±π / 2ρ(θ0)

[0009] S2: When t=1,2,…,T, the loop executes the following steps:

[0010] S21: Quantum Server for Parameterized Quantum States Constructing a classic shadow tomography SG t ,in Represents θ k is θ k ±v / 2ρ(θ t )

[0011] S22: When i=1,2,…,N, the following steps are executed cyclically:

[0012] S221: The i-th classical edge client downloads the classical shadow tomography SG from the quantum server t

[0013] S222: The i-th classic edge client according to SG t , data encoding observation O i And the cross entropy loss function L(θ t ) Calculate the local gradient parameter gL i,t

[0014] in

[0015] S223: The i-th classic edge client uploads the local gradient parameter gL after local training i,t

[0016] S224: size M of local training data uploaded by the i-th classic edge client i,t

[0017] S225: Increase the value of i by one

[0018] S23: The quantum server calculates the total amount of training data for this round

[0019] S24: Quantum server aggregates all local gradient parameters To obtain the global gradient parameter

[0020]

[0021] S25: Quantum server updates model parameters θ t

[0022] S26: Let the value of t increase by one

[0023] S3: When the preset number of iterations is reached, the training stops and the training model parameters θ are returned. T

[0024] A short video content classification method based on quantum federated learning, whose architecture is a server and multiple edge clients. The server is a quantum computer with quantum computing capabilities, and the edge clients only have classical computing capabilities. The server and edge clients communicate through classical channels.

[0025] The quantum server is used to aggregate the local gradient parameters uploaded by each classic edge client to calculate the global gradient parameters. The classic edge client has local training data for training the local model. Its local user data is not uploaded to the quantum server. Each classic edge client uses the same local training model.

[0026] The embodiments of the present invention have the following advantages:

[0027] 1. The "quantum server-classical client" architecture only requires quantum computing capabilities for the server, while the mobile edge devices only need to use ordinary classical computing devices. This not only effectively utilizes the computing advantages of quantum computing, but also does not require the user's equipment to be upgraded, achieving the effect of reducing costs and increasing efficiency.

[0028] 2. The quantum federated learning framework is a distributed machine learning framework. There is no need to exchange original data between the mobile edge device and the quantum server. Only the model training parameters and intermediate results need to be transferred. That is, the mobile edge device downloads the initial parameters from the quantum server through the classical channel, then trains the model with the local album data, and uploads the updated model parameters to the quantum server through the classical channel. The quantum server aggregates the parameters and then returns them to each mobile edge device to update the local model parameters. In this process, the user's local privacy data is not transferred, thus realizing the privacy protection function. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is a schematic diagram of the specific system structure provided by the present invention. DETAILED DESCRIPTION

[0030] The present invention is further described below in conjunction with the embodiments. People familiar with the technology can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0031] See attached Figure 1 The present invention provides a short video content classification method based on quantum federated learning, comprising:

[0032] S1: The quantum server initializes the model with random model parameters θ0 and independently measures each parameterized quantum state with a random Pauli matrix basis Repeat M times to construct a classical shadow tomography SG0, and each measurement output is a bit string of length n |b(θ)>, where

[0033]

[0034] Represents θ k is θ k ±π / 2ρ(θ0)

[0035] S2: When t=1,2,…,T, the loop executes the following steps:

[0036] S21: Quantum server independently measures each parameterized quantum state using a random Pauli matrix basis Repeat M times to construct a classical shadow tomography SG t , each measurement output is a bit string of length n |b(θ)>, where Represents θ k is θ k ±π / 2ρ(θ t )

[0037] S22: When i=1,2,…,N, the following steps are executed cyclically:

[0038] S221: The i-th classical edge client downloads the classical shadow tomography SG from the quantum server t

[0039] S222: The i-th classic edge client adds s based on the local model and the album z unlabeled image data to obtain the client's interest vector v for each type of short video content i =[v1,…,v N ] This vector is then encoded as an observable in is the Pauli matrix Z acting on the jth quantum bit. Then, according to the classical shadow tomography SG downloaded from the quantum server t , its own data encoding observation O i (v) and the cross entropy loss function L(θ t ) Calculate the local gradient parameters

[0040] S223: The i-th classic edge client uploads the local gradient parameter gL after local training i,t

[0041] S224: Size Mi of local training data uploaded by the i-th classic edge client ,t

[0042] S225: Increase the value of i by one

[0043] S23: The quantum server calculates the total amount of training data for this round

[0044] S24: Quantum server aggregates all local gradient parameters To obtain the global gradient parameter

[0045] And calculate the interest vector of the user group

[0046] S25: Quantum server updates model parameters θ t

[0047] S26: Let the value of t increase by one

[0048] S3: When the preset number of iterations is reached, the training stops and the training model parameters θ are returned. T

[0049] The present invention provides a short video content classification method based on quantum federated learning, and its architecture is a server and multiple edge clients. The server is a quantum computer with quantum computing capabilities, and the edge client only has classical computing capabilities. The server and the edge client communicate through a classical channel.

[0050] The quantum server is used to aggregate the local gradient parameters uploaded by each classic edge client to calculate the global gradient parameters. The classic edge client has local training data for training the local model. Its local user data is not uploaded to the quantum server. Each classic edge client uses the same local training model.

[0051] The above is only a preferred embodiment of the present invention. Any person skilled in the art may modify the present invention by using the above technical solution or modify it into an equivalent technical solution. Therefore, any simple modification or equivalent replacement made according to the technical solution of the present invention shall fall within the scope of protection claimed by the present invention.

Claims

1. A short video content classification method based on quantum federated learning, characterized in that: include: S1: The quantum server initializes the model with random model parameters θ0 and performs parameterized quantum state Construct a classic shadow tomography SG0, where Represents θ k is θ k ±π / 2ρ(θ0), V(θ) is a parameterized quantum circuit S2: When t=1,2,…,T, the loop executes the following steps: S21: Quantum Server for Parameterized Quantum States Constructing a classic shadow tomography SG t ,in Represents θ k is θ k ±π / 2ρ(θ t ) S22: When i=1,2,…,N, the following steps are executed cyclically: S221: The i-th classical edge client downloads the classical shadow tomography SG from the quantum server t S222: The i-th classic edge client according to SG t , data encoding observation O i And the cross entropy loss function L(θ t ) Calculate the local gradient parameter gL i,t in S223: The i-th classic edge client uploads the local gradient parameter gL after local training i,t S224: The size M of the local training data of the tth round uploaded by the i-th classic edge client i,t S225: Increase the value of i by one S23: The quantum server calculates the total amount of training data for this round S24: Quantum server aggregates all local gradient parameters To obtain the global gradient parameter S25: Quantum server updates model parameters θ t S26: Let the value of t increase by one S3: When the preset number of iterations is reached, the training stops and the training model parameters θ are returned. T 2. According to claim 1, a short video content classification method based on quantum federated learning is characterized in that: The server is a quantum computer with quantum computing capabilities, and the edge client only has classical computing capabilities. The server and edge client communicate through classical channels. The number of servers is 1 and the number of edge clients is multiple. The quantum server is used to aggregate the local gradient parameters uploaded by each classic edge client to calculate the global gradient parameters. The classic edge client has local training data for training the local model. Its local user data is not uploaded to the quantum server. Each classic edge client uses the same local training model.

Images