DETAILED ACTION
This action is written in response to the RCE filed 6/1/26. The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Response to Arguments
The Applicants argue that the previous art of record does not anticipate or render obvious the claims as currently amended. The Examiner provides updated prior art rejections below necessitated by the current amendments.
Subject Matter Eligibility
In determining whether the claims are subject matter eligible, the examiner has considered and applied the 2019 USPTO Patent Eligibility Guidelines, as well as guidance in the MPEP chapter 2106. The examiner finds that the independent claims are directed to the practical application of improving federated learning via hierarchical aggregation.
Claim Rejections - 35 USC § 112(a) - Written Description
The following is a quotation of the first paragraph of 35 U.S.C. 112(a):
(a) IN GENERAL.—The specification shall contain a written description of the invention, and of the manner and process of making and using it, in such full, clear, concise, and exact terms as to enable any person skilled in the art to which it pertains, or with which it is most nearly connected, to make and use the same, and shall set forth the best mode contemplated by the inventor or joint inventor of carrying out the invention.
Claims 1-20 are rejected under 35 U.S.C. 112(a) as failing to comply with the written description requirement. The claims contains subject matter which was not described in the specification in such a way as to reasonably convey to one skilled in the relevant art that the inventors, at the time the application was filed, had possession of the claimed invention. Claim 1 recites, in part, “receiving an initial response to the query from the first party that does not include the first party local data or a first party local model derived from the first party local data;” (Emphasis added.) The examiner was unable to identify support in the Applicant’s originally filed specification for this particular negative limitation. Therefore, this limitation is new matter.
For the above reasons, claim 1 is rejected as failing to meet the written description requirement. This rejection applies equally to independent claims 8 and 15, as well as to all pending dependent claims.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103(a) which forms the basis for all obviousness rejections set forth in this Office action:
(a) A patent may not be obtained though the invention is not identically disclosed or described as set forth in section 102 of this title, if the differences between the subject matter sought to be patented and the prior art are such that the subject matter as a whole would have been obvious at the time the invention was made to a person having ordinary skill in the art to which said subject matter pertains. Patentability shall not be negatived by the manner in which the invention was made.
The following are the references relied upon in the rejections below:
Chai (Chai, Zheng, et al. "Tifl: A tier-based federated learning system." Proceedings of the 29th international symposium on high-performance parallel and distributed computing. 2020. Cited by Applicant in IDS dated 6/21/22.)
McMahan (McMahan, Brendan, et al. "Communication-efficient learning of deep networks from decentralized data." Artificial intelligence and statistics. Pmlr, 2017.)
Verbraeken (Verbraeken, Joost, Matthijs Wolting, Jonathan Katzy, Jeroen Kloppenburg, Tim Verbelen, and Jan S. Rellermeyer. "A survey on distributed machine learning." Acm computing surveys (csur) 53, no. 2 (2020): 1-33.)
Zhang (Zhang, Tao, et al. "Privacy-preserving asynchronous grouped federated learning for IoT." IEEE Internet of Things Journal 9.7 (2021): 5511-5523.)
Claims 1-2, 4, 6-9, 11, 13-16, 18 and 20 are rejected under 35 U.S.C. 103(a) as being obvious over Chai and McMahan.
Regarding claims 1, 8 and 15, Chai discloses a processor-implemented method, (and a related system and computer program product) the method comprising:
initializing a plurality of aggregation groups including a plurality of parties and a plurality of local aggregators, the parties having access to stored local data; …
PNG
media_image1.png
340
422
media_image1.png
Greyscale
P. 128, fig. 2.
‘Aggregation groups’ :: tiers
‘Plurality of parties’ :: clients
‘Local aggregators’ :: P. 128, second col. “It is worth to note that in Fig. 2, we only show a single aggregator rather than the hierarchical master-child aggregator design for a clean presentation purpose. For large scale system in practice, TiFL supports master-child aggregator design for scalability and fault tolerance.”
P. 1, introduction, “Federated Learning (FL) [15] shines light on a new emerging high performance computing paradigm by addressing the security and privacy challenges through utilizing decentralized data that is training local models on the local data of each client (data parties)”. (Emphasis added.)
submitting the initial response to a first local aggregator from the plurality of local aggregators;
See fig. 2: clients [Wingdings font/0xE0] tiered aggregation groups [Wingdings font/0xE0] (global) aggregator.
submitting a final response from the first local aggregator or a second local aggregator from the plurality of local aggregators to a global aggregator; and
See fig. 2; the depicted ‘Aggregator’ is a global aggregator which receives updates Wi from each local aggregator.
building a machine learning model based on the final response.
Id.
McMahan discloses the following further limitation which Chai does not disclose:
submitting a query to a first party from the plurality of parties, the first party having access to first party local data that is stored on at least one device and accessible only to the first party;
P. 1, sec. 1, “Each client has a local training dataset which is never uploaded to the server. Instead, each client computes an update to the current global model maintained by the server, and only this update is communicated.”
P. 5, algorithm 1 (reproduced below).
PNG
media_image2.png
434
462
media_image2.png
Greyscale
‘query’ :: The examiner interprets this term according to its broadest reasonable interpretation as encompassing the following step in McMahan: The server initiates each round (epoch) by communicating weight values (initialized or updated) to each client. Each client responds by communicating updated weight values based on local training.
receiving an initial response to the query from the first party that is based on the first party local data, the initial response not including the first party local data or a first party local model derived from the first party local data;
P. 5, algorithm 1. The examiner interprets “the initial response not including … a first party local model” as encompassing communicating weight values only from the local client to the server. This claim limitation is a negative limitation. Because the information communicated does not include neural network architecture, activation parameters, or other model parameters, it does not comprise an entire model. In other words, model weights only are not a “local model”, but merely a part of a local model.
At the time of filing, it would have been obvious to a person of ordinary skill to apply the federated learning techniques disclosed by McMahan with the system of Chai because this would provide for effective model training while preserving data priviacy with respect to local data, as well as reducing communication costs that would be associated with training data transfer. Both disclosures pertain to federated machine learning.
Regarding independent claims 8 and 15, the computer hardware components recited therein (ie “one or more processors, one or more computer-readable memories, [and] one or more computer-readable tangible storage medium” and “one or more computer-readable tangible storage medium”) are inherent throughout the Chai disclosure.
Regarding claims 2, 9 and 16, Chai discloses the further limitation comprising:
submitting an intermediary response from the first local aggregator or a first intermediary aggregator to the second local aggregator, the first intermediary aggregator, or a second intermediary local aggregator.
P. 131, “[4] introduces multiple levels of server aggregators in order to achieve scalability and fault tolerance in extreme scale situations, i.e., with millions of clients.”
Regarding claims 4, 11 and 18, Chai discloses the further limitation wherein the plurality of aggregation groups each correspond to a physical location.
P. 125, “In conventional high-performance computing (HPC), all the data is collected and centralized in one location and proceed by supercomputers with hundreds to thousands of computing nodes. However, security and privacy concerns have led to new legislation such as the General Data Protection Regulation (GDPR) [27] and the Health Insurance Portability and Accountability Act (HIPAA) [24] that prevent transmitting data to a centralized location, thus making conventional high performance computing difficult to be applied for collecting and processing the decentralized data. Federated Learning (FL) [15] shines light on a new emerging high performance computing paradigm by addressing the security and privacy challenges through utilizing decentralized data that is training local models on the local data of each client (data parties) and using a central aggregator to accumulate the learned gradients of local models to train a global model.” (Emphasis added.)
Regarding claims 6, 13 and 20, Chai discloses the further limitation wherein the submitting further comprises:
submitting a plurality of initial responses to more than one local aggregator from the plurality of local aggregators.
PP. 128-29, “Different from vanilla FL that employs a random client selection policy, in TiFL the scheduler selects a tier and then randomly selects targeted number of clients from that tier. After the selection of clients, the training proceeds as state-of-the-art FL system does.”
P. 127, first col. “The vanilla FL algorithm is briefly summarized in Alg. 1. The aggregator first randomly initializes weights of the global model denoted by 𝜔0. At the beginning of each round, the aggregator sends the current model weights to a subset of randomly selected clients. Each selected client then trains its local model with its local data and sends back the updated weights to the aggregator after local training. At each round, the aggregator waits until all selected clients respond with their corresponding trained weights. This iterative process keeps on updating the global model until a certain number of rounds are completed or a desired accuracy is reached.”
Regarding claims 7 and 14, Chai discloses the further limitation wherein each party that submits a plurality of responses submits responses to different local aggregators so each local aggregator receives an incomplete subset of the plurality of responses.
P. 128, fig. 2 (reproduced supra), illustrating a hierarchical aggregation scheme. As described more fully in sec. 4.1, each ‘tier’ comprises at least one local aggregator, and worker nodes receive updates only from their immediate parent aggregator, and not from all aggregators in the system.
Claims 3, 5, 10, 12, 17 and 19 are rejected under 35 U.S.C. 103 as being unpatentable over Chai, McMahan and Zhang.
Regarding claims 3, 10 and 17, Zhang discloses the following further limitation which Chai/McMahan do not disclose wherein each local aggregator from the plurality of local aggregators selects an aggregation method, and wherein the final response is determined using the aggregation method.
P. 5513, “C. Our Innovation To avoid the shortcomings of DP-based methods and cryptography-based methods in the existing FL for IoT, RDP is applied into our framework to enable strong composition privacy results. We design an adaptive RDP-based privacy budget allocation protocol since the direct using of RDP cannot enable the models’ convergence to be an optimal one. Our framework enables the server to adjust the privacy budget of the corresponding model dynamically according to the accuracy of each local model on the public validation data set, which can tradeoff the utility of the global model and privacy guarantee. Next, the local model is perturbed according to the allocated privacy budget in the local training phase.”
At the time of filing, it would have been obvious to a person of ordinary skill to apply the Zhang technique for aggregator-based differential privacy within a federated learning system to the local aggregators of Zhang/McMahan because this would provide for enhanced performance in the face of heterogeneous (and possibly malicious) worker nodes. Both Zhang and Chai pertain to federated learning.
Regarding claims 5, 12 and 19, Zhang discloses the following further limitation which Chai/McMahan do not disclose wherein a party from the plurality of parties can be removed from a first aggregation group of the plurality of aggregation groups or placed in a second aggregation group of the plurality of aggregation groups after the initializing of the plurality of aggregation groups.
PP. 5512-13, sec. A(1), “Adaptive Weight-Based Worker Selection in AFL”.
The obviousness analysis of claims 3/10/17 applies equally here.
Additional Relevant Prior Art
The following references were identified by the Examiner as being relevant to the disclosed invention, but are not relied upon in any particular prior art rejection:
Liu discloses a system for asynchronous federated learning in the presences of data heterogeneity. (US 2022/0383198 A1)
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to Vincent Gonzales whose telephone number is (571) 270-3837. The examiner can normally be reached on Monday-Friday 7 a.m. to 4 p.m. MT. If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Miranda Huang, can be reached at (571) 270-7092.
Information regarding the status of an application may be obtained from the USPTO Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov.
/Vincent Gonzales/Primary Examiner, Art Unit 2124
Prosecution Insights