Online learning method with delayed publication satisfying differential privacy
Through the method of delayed release and noise-added operation, the problem of large errors in the existing online learning algorithm is solved, and high-precision protection with low privacy losses and low communication overhead is achieved, which is suitable for online learning scenarios with privacy protection requirements.
Patent Information
- Application Number
- CN202110763046.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-07-06
- Publication Date
- 2024-12-10
- Estimated Expiration
- 2041-07-06
AI Technical Summary
The existing online learning algorithms that satisfy differential privacy have large errors and large communication overhead, making it difficult to maintain high accuracy while protecting privacy.
Using the method of delayed publishing, through delayed publishing and noise addition operations, select privacy parameters and publishing time, and use adaptive model update methods such as gradient method, Follow the Leader algorithm and Hedge algorithm to reduce the number of publishing rounds, only in Update and publish models if necessary.
It reduces privacy losses and communication overhead, improves the accuracy of online learning, and is suitable for scenarios with privacy protection needs.
Smart Images
Figure CN113379069B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the fields of privacy protection and machine learning, and in particular to an online learning method with delayed publishing that satisfies differential privacy. Background Art
[0002] Online learning, as an important branch of machine learning, is characterized by continuously updating the training model based on the user's online feedback information, and using the model to provide services to subsequent users. It can be seen that if the user's feedback information contains their private information, the training process may leak their privacy to subsequent users through the model. For online learning that protects privacy, an important research idea is to use differential privacy methods for design, that is, to introduce uncertainty into the training process through perturbation (noise), so that attackers cannot accurately infer the user's private information through the model. For online learning algorithms that meet differential privacy, the existing designs mainly include two ideas: 1) directly perturb the training model; 2) perturb some intermediate results in the training process. On the other hand, online learning algorithms that meet differential privacy can be extended to distributed scenarios.
[0003] The error of existing online algorithms that satisfy differential privacy (measured by regret loss, that is, the difference between the cumulative error of the model and the corresponding error of the optimal single model) is mostly far inferior to the regret loss of non-private online learning. Summary of the invention
[0004] In view of this, the main purpose of the present invention is to provide an online learning method with delayed release that satisfies differential privacy, in order to partially solve at least one of the above technical problems.
[0005] In order to achieve the above objectives, as one aspect of the present invention, an online learning method with delayed release and differential privacy is provided, which realizes online learning with high precision and privacy protection through delayed release and noise addition operations, including:
[0006] Select the privacy parameter (∈,δ) and release time 1≤t1 <t2<…<t m ≤T;
[0007] Randomly initialize and release the initial model x1;
[0008] At time t=1, 2, ..., T, update x according to certain rules t-1 get If t∈{t1,t2,…,t m}, then according to the privacy parameter (∈,δ) Add noise and publish the noised model x t .
[0009] Among them, it also includes if t is in t i to i+1 The model released by the learning party is x ti .
[0010] Among them, according to the specific online learning problem, in addition to privacy parameters and release time, the model feasible domain must also be selected At the same time, the release time can be selected in an adaptive way instead of in advance; if the model feasible domain is selected The initial model must fall within the model feasible domain middle.
[0011] Among them, the certain rule is an online learning model updating method, and the online learning model updating method includes a gradient method, a Follow the Leader algorithm and a Hedge algorithm.
[0012] Wherein, the updating of x according to certain rules t-1 get During the update process, if the update rules need to use the previous model, the noisy model or the model before noise can be used according to actual needs and algorithm design; if the model is not updated, the learning party can choose not to publish the model, and the user can directly call the last published model to reduce communication overhead.
[0013] As another aspect of the present invention, a distributed online learning method with delayed release and differential privacy is provided, which realizes distributed online learning with high precision and privacy protection through delayed release and noise addition operations, including:
[0014] All learners select privacy parameters (∈,δ) and release time 1≤t1 <t2<…<t m ≤T; Each learning party i∈{1,2,…,M} randomly initializes and publishes the initial model
[0015] At time t=1,2,…,T, update x according to certain rules t-1 get If t∈{t1,t2,…,t m}, then according to the privacy parameter (∈,δ) Add noise and publish the noised model x t , and x t Broadcast to its neighbor learners.
[0016] Among them, it also includes if t is in t i to i+1 The model released by the learning party is x ti And broadcast it to neighbor learners.
[0017] Among them, according to the specific online learning problem, in addition to privacy parameters and release time, the model feasible domain must also be selected At the same time, the release time can be selected in an adaptive way instead of in advance; if the model feasible domain is selected The initial model must fall within the model feasible domain middle.
[0018] Among them, the certain rule is a distributed online learning model updating method, and the distributed online learning model updating method includes a gradient method.
[0019] Wherein, the updating of x according to certain rules t-1 get During the update process, if the model is not updated, the learning party does not need to publish the model, and the user can directly call the last published model to reduce communication overhead.
[0020] Based on the above technical solution, it can be seen that the online learning method with delayed release that satisfies differential privacy of the present invention has at least one of the following beneficial effects compared with the prior art:
[0021] 1) Since the number of release rounds is small, the corresponding privacy loss is small. By reasonably controlling the number of release rounds, a smaller regret loss can be obtained under the same privacy loss budget;
[0022] 2) The number of publishing rounds is small, so the communication overhead is small. BRIEF DESCRIPTION OF THE DRAWINGS
[0023] Figure 1 A schematic diagram of a flow chart of an online learning algorithm with delayed publishing and satisfying differential privacy provided in an embodiment of the present invention;
[0024] Figure 2 A schematic diagram of a distributed online learning algorithm flow with delayed publishing and satisfying differential privacy provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0025] The purpose of the present invention is to provide an online learning algorithm that satisfies differential privacy with delayed release, which can reduce the error of the online learning algorithm that satisfies differential privacy and has lower communication overhead.
[0026] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings.
[0027] like Figure 1 As shown, an embodiment of the present invention provides an online learning algorithm with delayed release that satisfies differential privacy. Through delayed release and noise addition operations, high-precision online learning with privacy protection is achieved. The method includes:
[0028] Step 1: Select the privacy parameter (∈,δ) and the release time 1≤t1 <t2<…<t m ≤T;
[0029] Step 2, randomly initialize and publish the initial model x1;
[0030] Step 3: At time t=1,2,…,T, update according to certain rules If t∈{t1,t2,…,t m}, then according to the privacy parameter (∈,δ) Add noise and publish the model after adding noise (denoted as x t ), if not, then publish
[0031] like Figure 2 As shown, the embodiment of the present invention provides a distributed online learning algorithm with delayed release and differential privacy. Through delayed release and noise addition operations, high-precision distributed online learning with privacy protection is achieved. The method includes:
[0032] Step 1: All learners select privacy parameters (∈,δ) and release time 1≤t1 <t2<…<t m ≤T;
[0033] Step 2: Each learning party i∈{1,2,…,M} randomly initializes and publishes the initial model
[0034] Step 3: At time t=1, 2, ..., T, update according to certain rules If t∈{t1,t2,…,t m}, then according to the privacy parameter (∈,δ) Add noise and publish the model after adding noise (denoted as x t ), and x t Broadcast to its neighbor learners, if not, publish And broadcast it to neighbor learners.
[0035] The method of the present invention achieves high-precision online learning with privacy protection through delayed release and noise addition. Due to the small number of release rounds, the present invention has the advantages of low privacy loss and low communication overhead, and has high accuracy under the same privacy loss budget, which is suitable for (distributed) online learning scenarios with privacy protection requirements.
[0036] The embodiments of the present invention are described in further detail below.
[0037] The online learning algorithm with delayed release and satisfying differential privacy provided by the embodiment of the present invention mainly includes the following steps:
[0038] Step 1: Select the privacy parameter (∈,δ) and the release time 1≤t1 <t2<…<t m ≤T;
[0039] Step 2, randomly initialize and publish the initial model x1;
[0040] Step 3: At time t=1,2,…,T, update according to certain rules If t∈{t1, t2, …, t m}, then according to the privacy parameter (∈,δ) Add noise and publish the model after adding noise (denoted as x t ), if not, then publish
[0041] The specific steps are as follows:
[0042] Step 1: For specific online learning problems, in addition to privacy parameters and release time, it may be necessary to select the model feasible domain At the same time, the release time does not need to be selected in advance, but can be selected using adaptive methods.
[0043] Step 2: Initialize the model to fall within the feasible domain of the model middle.
[0044] Step 3. Common online learning model update methods include gradient method, Follow the Leader algorithm, Hedge algorithm, etc. During the update process, if the update rule needs to use the previous model, the noisy model or the model before noise can be used according to actual needs and algorithm design. If the model is not updated, the learner can choose not to publish the model, and the user can directly call the last published model to reduce communication overhead.
[0045] The distributed online learning algorithm with delayed release and satisfying differential privacy provided by the embodiment of the present invention mainly includes the following steps:
[0046] Step 1: All learners select privacy parameters (∈,δ) and release time 1≤t1 <t2<…<t m ≤T;
[0047] Step 2: Each learning party i∈{1,2,…,M} randomly initializes and publishes the initial model
[0048] Step 3: At time t=1, 2, ..., T, update according to certain rules If t∈{t1,t2,…,t m}, then according to the privacy parameter (∈,δ) Add noise and publish the model after adding noise (denoted as x t ), and x t Broadcast to its neighbor learners, if not, publish And broadcast it to neighbor learners.
[0049] The specific steps are as follows:
[0050] Step 1: For specific online learning problems, in addition to privacy parameters and release time, it may be necessary to select the model feasible domain At the same time, the release time does not need to be selected in advance, but can be selected using adaptive methods.
[0051] Step 2: Initialize the model to fall within the feasible domain of the model middle.
[0052] Step 3: Common distributed online learning model update methods include gradient methods; if the model is not updated, the learning party may not publish the model, and the user can directly call the last published model to reduce communication overhead.
[0053] Due to the small number of publishing rounds, the present invention has the advantages of low privacy loss and low communication overhead, and has higher accuracy under the same privacy loss budget, and is suitable for (distributed) online learning scenarios with privacy protection requirements.
[0054] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. An online learning method with delayed release that satisfies differential privacy is suitable for distributed online learning scenarios with privacy protection requirements. Through delayed release and noise addition operations, high-precision online learning with privacy protection is achieved, including: Select the privacy parameter (∈,δ) and release time 1≤t1 <t2<…<t m ≤T; Randomly initialize and release the initial model x1; At time t=1,2,…,T, update x according to certain rules t-1 get If t∈{t1,t2,…,t m }, then according to the privacy parameter (∈,δ) Add noise and publish the noised model x t , the model x t It is trained based on user feedback information and released in the model x t There will be a need to protect feedback information during the process, where the feedback information includes privacy information; Depending on the specific online learning problem, in addition to privacy parameters and release time, the model feasible domain needs to be selected At the same time, the release time can be selected in an adaptive way instead of in advance; if the model feasible domain is selected The initial model must fall within the model feasible domain middle; The update x according to certain rules t-1 get During the update process, if the update rules need to use the previous model, the noisy model or the model before noise can be used according to actual needs and algorithm design; if the model is not updated, the learning party can choose not to publish the model, and the user can directly call the last published model to reduce communication overhead.
2. The online learning method according to claim 1, further comprising: i to i+1 The model released by the learning party is x ti .
3. According to the online learning method of claim 1, the certain rule is an online learning model updating method, and the online learning model updating method includes a gradient method, a Follow the Leader algorithm and a Hedge algorithm.
4. A distributed online learning method with delayed release and differential privacy, which is suitable for distributed online learning scenarios with privacy protection requirements. Through delayed release and noise addition operations, high-precision distributed online learning with privacy protection is achieved, including: All learners select privacy parameters (∈,δ) and release time 1≤t1 <t2<…<t m ≤T; Each learning party i∈{1,2,…,M} randomly initializes and publishes the initial model At time t=1,2,…,T, update x according to certain rules t-1 get If t∈{t1,t2,…,t m }, then according to the privacy parameter (∈,δ) Add noise and publish the noised model x t , and x t Broadcast to its neighboring learners, the model x t It is trained based on user feedback information and released in the model x t There will be a need to protect feedback information during the process, where the feedback information includes privacy information; Depending on the specific online learning problem, in addition to privacy parameters and release time, the model feasible domain needs to be selected At the same time, the release time can be selected in an adaptive way instead of in advance; if the model feasible domain is selected The initial model must fall within the model feasible domain middle; The update x according to certain rules t-1 get During the update process, if the model is not updated, the learning party does not need to publish the model, and the user can directly call the last published model to reduce communication overhead.
5. The distributed online learning method according to claim 4 further comprises: if t is t i to i+1 The model released by the learning party is x ti And broadcast it to neighbor learners.
6. The distributed online learning method according to claim 4, wherein the certain rule is a distributed online learning model updating method, and the distributed online learning model updating method comprises a gradient method.