Low-overhead federal zero-order fine tuning method and system for resource-constrained intelligent terminal

Abstract

The invention belongs to the technical field of artificial intelligence, and particularly relates to a low-overhead federal zero-order fine tuning method and system for a resource-constrained intelligent terminal, and the method comprises the steps: sharing a randomness mechanism through a cloud-side coordination server and a mobile intelligent terminal, and achieving the uploading of a small number of zero-order directional derivative scalar quantities at the mobile intelligent terminal; the server reconstructs a disturbance direction according to the shared random seed and forms a pseudo gradient, and performs coordinate-by-coordinate robust updating on an aggregation result through an adaptive reforming mechanism, so that efficient, stable and user-fair federal fine tuning is realized under the condition of not depending on explicit gradient uploading; two-point zero-order estimation is adopted by the mobile terminal, back propagation is not needed, disturbance directions are resampled on demand through shared random seeds, explicit storage of high-dimensional direction vectors and back propagation of intermediate states are not needed, and therefore memory overhead is reduced.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence technology, specifically relating to a low-overhead federated zero-order fine-tuning method and system for resource-constrained smart terminals. Background Technology

[0002] With the increasing prevalence of intelligent models deployed on mobile smart terminals such as smartphones, wearable devices, smart home terminals, and edge gateways, terminal-side models need continuous updates or lightweight fine-tuning based on local user data to adapt to different users' usage habits, interaction environments, and task requirements. In these application scenarios, user data is usually stored locally on each terminal and is not suitable for direct centralized uploading. Therefore, federated learning provides a feasible technical path for collaborative updates of terminal-side models. However, most existing federated learning methods rely on first-order gradient backpropagation and incremental uploading of high-dimensional gradients or models, which have the following shortcomings in resource-constrained mobile smart terminal scenarios: 1. High computational and memory overhead. Backpropagation requires saving intermediate activation and gradient information, and memory consumption during training is significantly higher than during inference. Under limitations of computing power and energy consumption, devices with weak computing power cannot sustain this for long periods. 2. High uplink communication pressure. Existing federated learning methods typically upload model increments, gradients, or compressed high-dimensional vectors, which are related to the dimensionality of the model parameters. As models with larger parameter sizes and multimodal models are increasingly applied to terminal collaboration scenarios, this has led to a continuous increase in the number of parameters. Simultaneously, the server-side many-to-one paradigm of federated learning can easily cause server-side bandwidth congestion when the number of clients increases, especially during peak communication information aggregation. 3. Heterogeneous data leads to training instability. Cross-device user data naturally exhibits heterogeneity, and multi-step local updates can easily cause client drift. Furthermore, zero-order estimation itself has high variance, further amplifying aggregation noise, leading to a decrease in average accuracy and performance degradation for tail users.

[0003] Furthermore, while existing zero-order optimization methods estimate the descent direction through function values, avoiding explicit backpropagation, their direct application in federated fine-tuning still presents several challenges: first, they still require transmitting high-dimensional update information, making it difficult to fundamentally reduce uplink communication overhead; second, they lack robust update mechanisms to address terminal heterogeneity, high variance in zero-order estimation, and the cumulative effect of aggregation noise, making it difficult to balance training stability with fairness among different users. Therefore, a federated model fine-tuning method suitable for resource-constrained mobile smart terminals is still needed to achieve low end-side overhead, low uplink communication, stable optimization in heterogeneous scenarios, and good user fairness without relying on explicit high-dimensional gradient uploads. Summary of the Invention

[0004] The purpose of this invention is to provide a low-overhead federated zero-order fine-tuning method for resource-constrained smart terminals, addressing the problems of high backpropagation overhead, large uplink communication volume, unstable training under non-independent and identically distributed data, and poor performance of tail users in existing federated fine-tuning methods in mobile smart terminal fine-tuning scenarios. The technical solution is as follows: A low-overhead federated zero-order fine-tuning method for resource-constrained smart terminals includes the following steps: S1: Initialization process: Initialize the global parameters related to the mobile smart terminal fine-tuning system; S2: Server broadcast process: The cloud-side coordination server broadcasts a random seed list and current global model parameters to the mobile smart terminals participating in this round of fine-tuning; S3: Gradient calculation process: Each mobile smart terminal performs two-point zero-order function estimation based on local data and server broadcast parameters, estimates the scalar update amount of the local working model, and completes the local lightweight parameter update based on the update amount; S4: Information transmission process: Each mobile smart terminal sends back the list of directional derivative scalars estimated in step S3 to the cloud-side coordination server. S5: Information Reconstruction Process: The cloud-side coordination server receives a list of directional derivative scalars from multiple mobile smart terminals, reconstructs the perturbation direction based on a shared random seed list, and aggregates them to form a pseudo gradient; S6: Robust server update process: The cloud side coordinates the server to maintain the first and second moments of the global update, and performs coordinate-level adaptive updates to the global model accordingly.

[0005] Preferably, the initialization parameters in step S1 include: Set global model parameters ;

[0006] Setting the global round number T controls the maximum number of global iterations, and setting the local round number... Control the number of local updates performed by the mobile smart terminal in each round of communication; Set terminal-side step size and server-side step size This is used to control the local update rate and the global update rate, and to initialize the first-order moment vector. Second-order moment vector .

[0007] Preferably, step S2 is as follows: S21. The cloud-side coordination server is in the... During round communication, the generated length is random seed sequence Each mobile smart terminal and the cloud-side coordination server can generate the same perturbation direction based on the same random seed using the consistent random direction generation function RANDDIRECTION; ; For the first Round Each local update step generates a random perturbation direction vector from the shared random seed according to the rule RANDDIRECTION. This direction vector should be a Gaussian random vector, a Rademacher random vector, or a random vector that satisfies the conditions of zero mean and unit variance.

[0008] S22. The cloud-side coordination server will select target terminals from the set of online mobile smart terminals to participate in this round of fine-tuning, and will then set the current global model parameters. With the generated random seed sequence Broadcast to the target terminal.

[0009] Preferably, step S3 is as follows: S31. Each mobile smart terminal receives the current global model parameters. With random seed sequence Then, initialize the local working model as follows: ; S32. For each local update step Mobile smart terminals based on random seeds Sampling random perturbation direction And sample small batches of data from local data. ; S33. The client calculates and stores the scalar values ​​of the zeroth-order directional derivatives at the two points: ; in It is a smoothing factor; Indicates the first Round communication The mobile smart terminal in the first Each local update step samples a small batch of data from its local dataset; Indicates the first Round communication The mobile smart terminal in the first The local working model parameters corresponding to each local update step.

[0010] S34. The mobile intelligent terminal performs local parameter updates based on the scalar value of the directional derivative: .

[0011] Preferably, step S4 is in After each local update, each mobile smart terminal transmits data to the cloud-based coordination server. scalar of directional derivative Transmission dimension It is completely decoupled from the number of model parameters.

[0012] Preferably, step S5 is as follows: S51. The cloud-side coordination server receives directional derivative scalars from multiple mobile smart terminals, for each local step. Calculate the average separately to obtain ; Indicates the first The number of mobile smart terminals involved in wheel-to-wheel operation.

[0013] S52. The cloud-based coordination server uses a random seed sequence. Reconstruct the corresponding perturbation direction ; S53. Cloud-based coordination server reconstructs aggregated pseudo-gradients. And store it on the server for subsequent updates.

[0014] Preferably, step S6 is as follows: S61. The cloud-side coordination server converts the aggregated pseudo-gradient described in S5 into a drift-type parameter update. And update the first-order moment based on this. Second-order moment information ,in, This represents the Hadamard element-wise product. This represents the attenuation coefficient of the first moment. This represents the attenuation coefficient of the second moment; S62. The cloud-based coordination server performs a coordinate-by-coordinate update on the global model based on the first and second moments: .

[0015] S63. Then proceed to the next round of server broadcasting until the preset number of communication rounds T is reached or the convergence condition is met.

[0016] A low-overhead federated zero-order fine-tuning system for resource-constrained smart terminals includes a local zero-order gradient estimation module, a shared stochasticity management module, a pseudo-gradient reconstruction module, and a server robust update module. The local zeroth-order gradient estimation module is set up on the mobile smart terminal and calculates the directional derivative scalar through two-point zeroth-order function estimation. The shared randomness management module is deployed on the cloud-side coordination server and mobile intelligent terminal. It ensures that both parties generate consistent perturbation directions by broadcasting random seed sequences, thereby decoupling communication overhead from model dimensions. The pseudo-gradient reconstruction module is set up on a cloud coordination server, which is responsible for receiving the scalar list returned by the mobile smart terminal and combining it with a random seed to restore the perturbation vector, thereby aggregating and forming a pseudo-gradient. The robust update module is set up on a cloud-based coordination server. By maintaining the first and second moment information of global updates, it performs coordinate-level updates using an adaptive renormalization mechanism to suppress variance and drift caused by data heterogeneity, thereby improving training stability and user fairness.

[0017] Preferably, each mobile smart terminal uploads only K directional derivative scalars in each round, and completes the local update by using two-point zero-order forward evaluation, without performing a complete backpropagation or caching the intermediate results of backpropagation.

[0018] Compared with the prior art, the beneficial effects of this application are as follows: 1. Significantly reduces uplink communication overhead. Each mobile smart terminal only uploads data per round. Instead of uploading high-dimensional gradient vectors, model differences, or intermediate activation information, the uplink communication complexity is decoupled from the model parameter dimension. This makes it particularly suitable for lightweight model update scenarios with a large number of terminals and limited bandwidth, such as mobile phones, wearable devices, smart home terminals, and edge gateways.

[0019] 2. Reduced training memory and computational burden on mobile devices. Mobile smart terminals use two-point zero-order forward evaluation to complete local updates, eliminating the need for full backpropagation or caching intermediate backpropagation results. This significantly reduces peak memory usage and computational complexity, making it more suitable for fine-tuning online or quasi-online models on resource-constrained devices.

[0020] 3. Improve optimization stability under heterogeneous data. The server performs adaptive renormalization on the reconstructed pseudo-gradient, which can suppress long-term high-variance coordinates and extreme client signals, and alleviate drift caused by non-independent and identically distributed data and multi-step local updates.

[0021] 4. Improve user fairness and tail performance. By maintaining the second moment, it is possible to reduce the update contribution of high-variance users, reduce the dominance of a few high-amplitude update terminals on the global model, and improve the model performance on terminals corresponding to weak users or long-tail users. Attached Figure Description

[0022] Figure 1 This is a schematic diagram illustrating the stages of the present invention; Figure 2 This is a schematic diagram illustrating the specific process of the present invention. Detailed Implementation

[0023] To enable those skilled in the art to better understand the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] This invention provides a low-communication-overhead federated zero-order model fine-tuning method for resource-constrained smart terminals, applicable to lightweight model update scenarios on mobile smart terminals such as smartphones, wearable devices, smart home terminals, and edge gateways. Figure 1 As shown, this method utilizes a shared randomness mechanism between a cloud-side coordination server and multiple mobile smart terminals, enabling the terminal side to perform only two-point zero-order function estimations and upload a small number of directional derivative scalars. The cloud-side coordination server reconstructs the perturbation direction based on the shared random seed and forms an aggregated pseudo-gradient. Then, combined with an adaptive renormalization mechanism, the aggregated result is robustly updated coordinate-by-coordinate. This achieves stable federated fine-tuning under conditions of low communication overhead, low terminal resource consumption, and heterogeneous data without relying on explicit high-dimensional gradient uploads and complete backpropagation. Figure 2 As shown, the specific implementation method is as follows:

[0025] (1) Initialization phase: (1.1) Set the initial values ​​of the model parameters The initial values ​​can be initialized randomly, with zero initialization, or with pre-trained parameters. Preferably, the model parameters are a subset of trainable parameters in the model to be fine-tuned. The global model can be a fully parameter-trainable model or a parameter-efficient fine-tuning model, for example, training only some adapter parameters, low-rank matrix parameters, classification head parameters, or specified submodule parameters while freezing the remaining backbone parameters; at the same time, this invention is compatible with parameter-efficient fine-tuning methods such as LoRA.

[0026] (1.2) Set the global round T to control the maximum number of global iterations, and set the local round K to control the number of local updates of the mobile smart terminal in each round of communication.

[0027] (1.3) Set the terminal-side step size and server-side step size This is used to control the local update rate and the global update rate, and to initialize the first-order moment vector. Second-order moment vector .

[0028] (1.4) Set the smoothing factor Used for zero-order gradient estimation, setting adaptive renormalization parameters. and stable terms On an instance, it can be set However, this invention is not limited to the above values ​​and can be adjusted according to the model size, terminal resource conditions, and task requirements.

[0029] (2) Broadcasting stage: (2.1) In At the start of round-robin communication, the server generates a set of shared random seed sequences. Each mobile smart terminal and the cloud-side coordination server can generate the same perturbation direction based on the same random seed using a consistent random direction generation function RANDDIRECTION. .

[0030] (2.2) The cloud-side coordination server will coordinate the current model The generated random seed sequence is broadcast to the mobile smart terminals participating in this round of federal fine-tuning.

[0031] (3) Terminal local zero-level update stage: (3.1) Each mobile smart terminal Upon receiving the broadcast from the cloud-side coordination server After obtaining the random seed sequence, initialize the local working model as follows: .

[0032] (3.2) For each local update step Mobile smart terminals based on random seeds Sampling random perturbation direction The perturbation direction is preferably a Gaussian random direction or a random direction that satisfies the conditions of zero mean and unit variance.

[0033] (3.3) Mobile smart terminals sample small batches of data The sampling method can be uniform random sampling, stratified sampling, class-balanced sampling, or sliding window sampling.

[0034] (3.4) The mobile intelligent terminal is based on the current local working model parameters Smooth radius and the direction of the disturbance Calculate the loss function values ​​at the positive and negative perturbation points respectively: ; in, This indicates that the i-th mobile smart terminal is in the current small batch. The empirical loss function on.

[0035] (3.5) The mobile intelligent terminal calculates the scalar directional derivative estimate based on the above two function values: ; in As a smoothing factor, the initialization process is shared globally.

[0036] (3.6) The direction obtained by the reconstruction of the mobile intelligent terminal and scalar Perform local parameter updates: .

[0037] (4) Information transmission stage: (4.1) In After each local step update, each mobile smart terminal transmits K directional derivative scalars to the cloud-side coordination server. Transmission dimension It is completely decoupled from the number of model parameters.

[0038] (5) Information reconstruction stage: (5.1) After collecting all scalar sequences uploaded by participating mobile smart terminals in this round, the cloud-side coordination server indexes each local step. Statistical analysis was conducted separately. Specifically, the data uploaded by all terminals was analyzed first. Calculate the average of the scalars: .

[0039] (5.2) The cloud-side coordination server coordinates according to the random seed sequence Reconstruct the corresponding perturbation direction .

[0040] (5.3) The cloud-side coordination server forms an aggregated pseudo-gradient based on the average directional derivative scalar and the reconstructed perturbation direction: And store it on the server for subsequent updates.

[0041] (6) Server update phase: (6.1) The cloud-side coordination server converts the aggregated pseudo-gradient into drift-type parameter update. And update the first-order moment based on this. Second-order moment information .in, This represents the element-wise product of Hadamard.

[0042] (6.2) The cloud-side coordination server performs coordinate-by-coordinate updates on the global model based on the first and second moments: .

[0043] (6.3) After the global model update is completed, the cloud-side coordination server enters the next round of broadcasting until the preset number of communication rounds is reached. Alternatively, it may meet preset convergence conditions. These convergence conditions may include one or more of the following: the change in validation loss is less than a threshold, the improvement in model performance is less than a threshold, or the magnitude of global parameter updates is less than a threshold.

[0044] A low-overhead federated zero-order fine-tuning system for resource-constrained smart terminals is disclosed. The system comprises a cloud-side coordination server and multiple mobile smart terminals. The fine-tuning system includes: a local zero-order gradient estimation module, a shared stochasticity management module, a pseudo-gradient reconstruction module, and a server robust update module. The local zero-order gradient estimation module is located on the mobile smart terminal and calculates the directional derivative scalar through two-point zero-order function estimation, achieving lightweight local parameter updates without backpropagation to reduce terminal computational overhead. The shared stochasticity management module is deployed between the cloud-side coordination server and the mobile smart terminals, ensuring consistent perturbation directions through broadcasting random seed sequences, decoupling communication overhead from model dimensions. The pseudo-gradient reconstruction module is located on the cloud-side coordination server and receives the scalar list returned by the mobile smart terminals, reconstructing the perturbation vector using random seeds to aggregate and form pseudo-gradients. The server robust update module is located on the cloud-side coordination server and maintains globally updated first and second moment information, performing coordinate-level updates using an adaptive renormalization mechanism to suppress variance and drift caused by data heterogeneity, improving training stability and user fairness.

[0045] The present invention also discloses an apparatus including a processor and a memory; wherein the processor performs the steps of a low-overhead federated zero-order fine-tuning method for resource-constrained smart terminals.

[0046] In some embodiments, the processor may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory or process data, such as executing the multimodal emotion recognition method based on hypergraph diffusion and evidence fusion.

[0047] In some embodiments, the memory may be an internal storage unit of the terminal, such as a hard drive or RAM. In other embodiments, the memory may be an external storage device of the terminal, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the memory may include both internal and external storage units. The memory is used to store application software and various types of data installed on the terminal, such as the program code installed on the terminal. The memory can also be used to temporarily store data that has been output or will be output.

[0048] The display is used to show information on the terminal and to display a visual user interface. The components of the terminal communicate with each other via a system bus.

[0049] The present invention also provides a computer-readable storage medium storing a program for a low-overhead federated zero-order fine-tuning method for resource-constrained smart terminals, which, when executed by a processor, implements the steps of the low-overhead federated zero-order fine-tuning method for resource-constrained smart terminals as described above.

[0050] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.). The program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The computer-readable storage medium can be a memory, magnetic disk, optical disk, etc.

[0051] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A low-overhead federated zero-order fine-tuning method for resource-constrained smart terminals, characterized in that, Includes the following steps: S1: Initialization process: Initialize the global parameters related to the mobile smart terminal fine-tuning system; S2: Server broadcast process: The cloud-side coordination server broadcasts a random seed list and current global model parameters to the mobile smart terminals participating in this round of fine-tuning; S3: Gradient calculation process: Each mobile smart terminal performs two-point zero-order function estimation based on local data and server broadcast parameters, estimates the scalar update amount of the local working model, and completes the local lightweight parameter update based on the update amount; S4: Information transmission process: Each mobile smart terminal sends back the list of directional derivative scalars estimated in step S3 to the cloud-side coordination server. S5: Information Reconstruction Process: The cloud-side coordination server receives a list of directional derivative scalars from multiple mobile smart terminals, reconstructs the perturbation direction based on a shared random seed list, and aggregates them to form a pseudo gradient; S6: Robust server update process: The cloud side coordinates the server to maintain the first and second moments of the global update, and performs coordinate-level adaptive updates to the global model accordingly.

2. The low-overhead federated zero-order fine-tuning method for resource-constrained smart terminals according to claim 1, characterized in that, Step S1 initializes the parameters, including: Set global model parameters ; Setting the global round number T controls the maximum number of global iterations, and setting the local round number... Control the number of local updates performed by the mobile smart terminal in each round of communication; Set terminal-side step size and server-side step size This is used to control the local update rate and the global update rate, and to initialize the first-order moment vector. Second-order moment vector .

3. The low-overhead federated zero-order fine-tuning method for resource-constrained smart terminals according to claim 1, characterized in that, Step S2 is as follows: S21. The cloud-side coordination server is in the... During round communication, the generated length is random seed sequence Each mobile smart terminal and the cloud-side coordination server can generate the same perturbation direction based on the same random seed using the consistent random direction generation function RANDDIRECTION; ； For the first Round Each local update step is generated by a random perturbation direction vector from the shared random seed according to the rule RANDDIRECTION; S22. The cloud-side coordination server will select target terminals from the set of online mobile smart terminals to participate in this round of fine-tuning, and will then set the current global model parameters. With the generated random seed sequence Broadcast to the target terminal.

4. The low-overhead federated zero-order fine-tuning method for resource-constrained smart terminals according to claim 2, characterized in that, Step S3 is as follows: S31. Each mobile smart terminal receives the current global model parameters. With random seed sequence Then, initialize the local working model as follows: ; S32. For each local update step Mobile smart terminals based on random seeds Sampling random perturbation direction And sample small batches of data from local data. ; S33. The client calculates and stores the scalar values ​​of the zeroth-order directional derivatives at the two points: ； ； ； in, Indicates the first A small batch of mobile smart terminals The empirical loss function on; It is a smoothing factor; Indicates the first Round communication The mobile smart terminal in the first Each local update step samples a small batch of data from its local dataset; Indicates the first Round communication The mobile smart terminal in the first The local working model parameters corresponding to each local update step; S34. The mobile intelligent terminal performs local parameter updates based on the scalar value of the directional derivative: 。 5. The low-overhead federated zero-order fine-tuning method for resource-constrained smart terminals according to claim 2, characterized in that, Step S4 in After each local step update, each mobile smart terminal transmits K directional derivative scalars to the cloud-side coordination server. Transmission dimension It is completely decoupled from the number of model parameters.

6. The low-overhead federated zero-order fine-tuning method for resource-constrained smart terminals according to claim 2, characterized in that, Step S5 is as follows: S51. The cloud-side coordination server receives directional derivative scalars from multiple mobile smart terminals, for each local step. Calculate the average separately to obtain ; Indicates the first The number of mobile smart terminals involved in wheel-to-wheel operation; S52. The cloud-based coordination server uses a random seed sequence. Reconstruct the corresponding perturbation direction ; S53. Cloud-based coordination server reconstructs aggregated pseudo-gradients. , And store it on the server for subsequent updates.

7. The low-overhead federated zero-order fine-tuning method for resource-constrained smart terminals according to claim 6, characterized in that, Step S6 is as follows: S61. The cloud-based coordination server converts the aggregated pseudo-gradient described in S5 into a drift-type parameter update. And update the first-order moment based on this. Second-order moment information ,in, This represents the Hadamard element-wise product. This represents the attenuation coefficient of the first moment. This represents the attenuation coefficient of the second moment; S62. The cloud-based coordination server performs a coordinate-by-coordinate update on the global model based on the first and second moments: ； It is a stable term; S63. Then proceed to the next round of server broadcasting until the preset number of communication rounds T is reached or the convergence condition is met.

8. A low-overhead federated zero-order fine-tuning system for resource-constrained smart terminals, characterized in that, It includes a local zero-order gradient estimation module, a shared stochasticity management module, a pseudo-gradient reconstruction module, and a server robust update module; The local zeroth-order gradient estimation module is set up on the mobile smart terminal and calculates the directional derivative scalar through two-point zeroth-order function estimation. The shared randomness management module is deployed on the cloud-side coordination server and mobile smart terminal. It ensures that both parties generate consistent perturbation directions by broadcasting random seed sequences, thus decoupling communication overhead from model dimensions. The pseudo-gradient reconstruction module is set up on a cloud coordination server, which is responsible for receiving the scalar list returned by the mobile smart terminal and combining it with a random seed to restore the perturbation vector, thereby aggregating and forming a pseudo-gradient. The robust update module is set up on a cloud-based coordination server. It maintains the first and second moments of the global update and uses an adaptive remodeling mechanism to perform coordinate-level updates in order to suppress variance and drift caused by data heterogeneity.

9. The low-overhead federated zero-order fine-tuning system for resource-constrained smart terminals according to claim 8, characterized in that, Each mobile smart terminal uploads only K directional derivative scalars in each round, and completes local updates using two-point zero-order forward evaluation. There is no need to perform a complete backpropagation or cache intermediate backpropagation results.

Images