DETAILED ACTION
This action is responsive to the application filed on 11/24/2023. Claims 1-20 are pending in the case. Claims 1, 9, 11, and 14 are independent claims.
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Priority
Acknowledgment is made of applicant’s claim for domestic priority based on an international application, PCT/CN2022/093183, filed on 05/17/2022, which claims priority to a foreign application filed in People’s Republic of China, CN202110573525.4, on 05/25/2021. Receipt is acknowledged of certified copies of papers required by 37 CFR 1.55.
Information Disclosure Statement
The information disclosure statement filed 10/28/2024 fails to comply with 37 CFR 1.98(a)(3)(i) because it does not include a concise explanation of the relevance, as it is presently understood by the individual designated in 37 CFR 1.56(c) most knowledgeable about the content of the information, of each reference listed that is not in the English language. It has been placed in the application file, but the information referred to therein has not been considered.
The non-patent literature document “Chinese Office Action for Application No. 202110573525 dated October 21, 2023, 8 pages” has been stricken through and not considered. All other references are being considered by the examiner.
The information disclosure statement (IDS) submitted on 11/04/2024 is being considered by the examiner.
The information disclosure statement filed 01/09/2025 fails to comply with 37 CFR 1.98(a)(3)(i) because it does not include a concise explanation of the relevance, as it is presently understood by the individual designated in 37 CFR 1.56(c) most knowledgeable about the content of the information, of each reference listed that is not in the English language. It has been placed in the application file, but the information referred to therein has not been considered.
The non-patent literature document “International Search Report and Written Opinion Issued in PCT/CN2022/093183, dated July 22, 2022, 13 Pages” has been stricken through and not considered. All other references are being considered by the examiner.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1-10 and 14-20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Regarding claim 1:
Step 1 Statutory Category: Claim 1 is directed to a method, which falls under one of the four statutory categories.
Step 2A Prong 1 Judicial Exception: Claim 1 recites, in part, “training, …, the at least one first model to obtain at least one trained first model”. This limitation, under the broadest reasonable interpretation, and according to applicant’s specification paragraphs 0008, 0010, 0012, covers the recitation of mathematical concepts, see MPEP §2106.04(a)(2)(I). Further, the claim recites: “aggregating, …, the at least one trained first model to obtain an aggregation result; and updating, …, a locally stored second model to obtain an updated second model, wherein the updating is based on the aggregation result”. This limitation recites, under the broadest reasonable interpretation, covers the recitation of mathematical concepts, see MPEP §2106.04(a)(2)(I).
Step 2A Prong 2 Integration into a practical application: This judicial exception is not integrated into a practical application. In particular the claim recites: “A federated learning method”. This limitation is an additional element that generally links the use of the judicial exception to a particular technological environment or field of use. See MPEP §2106.05(h). Further, the claim recites: “receiving, …, information of at least one first model, wherein the information of the at least one first model is from at least one downstream device”. This limitation is an additional element that amounts to adding insignificant extra-solution activity to the judicial exception. See MPEP §2106.05(g). Further, the claim recites: “by a first server” and “by the first server”. These limitations are additional elements that amount to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer in its ordinary capacity as a tool to perform an existing process. See MPEP §2106.05(f).
Step 2B Significantly more: The claims do not include additional elements that are sufficient to amount to significantly more than the judicial exception. As discussed above with respect to integration of the abstract idea into a practical application, the additional elements: “A federated learning method” generally links the use of the judicial exception to a particular technological environment or field of use. Elements that merely generally link the use of the judicial exception to a particular technological environment or field of use cannot provide an inventive concept. Further, the additional element: “receiving, …, information of at least one first model, wherein the information of the at least one first model is from at least one downstream device” is insignificant extra-solution activity to the judicial exception, and further, is directed to receiving or transmitting data over a network which courts have recognized as well-understood, routine, and conventional when they are claimed in a generic manner, see MPEP §2106.05(d)(II). Further, the additional elements “by a first server” and “by the first server” amount to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer in its ordinary capacity as a tool to perform an existing process. Elements that merely amount to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer in its ordinary capacity as a tool to perform an existing process cannot provide an inventive concept. The claim is not patent eligible.
Regarding claim 2, the rejection of claim 1 is incorporated, and further, the claim recites: “training, …, the at least one first model under a constraint of a first sparsity constraint condition, to obtain the at least one trained first model, wherein the first sparsity constraint condition comprises a constraint formed by a function for reducing parameters in the first model”. This limitation recites mathematical concepts in addition to those identified in the rejection of the parent claim. Thus, the claim recites a judicial exception.
Further, the claim recites: “by the first server”. This limitation is an additional element that amounts to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer in its ordinary capacity as a tool to perform an existing process. See MPEP §2106.05(f). Elements that merely amount to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer in its ordinary capacity as a tool to perform an existing process cannot provide an inventive concept. The claim is not patent eligible.
Regarding claim 3, the rejection of claim 1 is incorporated, and further, the claim recites: “training, …, the at least one first model under constraints of a first sparsity constraint condition and a first distance constraint condition, to obtain the at least one trained first model, wherein the first sparsity constraint condition comprises a constraint formed by a function for reducing parameters in the first model, and the first distance constraint condition comprises a constraint formed by a similarity or a distance between the at least one first model and the second model”. This limitation recites mathematical concepts in addition to those identified in the rejection of the parent claim. Thus, the claim recites a judicial exception.
Further, the claim recites: “by the first server”. This limitation is an additional element that amounts to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer in its ordinary capacity as a tool to perform an existing process. See MPEP §2106.05(f). Elements that merely amount to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer in its ordinary capacity as a tool to perform an existing process cannot provide an inventive concept. The claim is not patent eligible.
Regarding claim 4, the rejection of claim 3 is incorporated, and further, the claim recites: “perturbing, …, the at least one first model by using noise, to obtain at least one updated first model”. This limitation recites mathematical concepts in addition to those identified in the rejection of the parent claim. Thus, the claim recites a judicial exception.
Further, the claim recites: “training, …, the at least one updated first model under the constraints of the first sparsity constraint condition and the first distance constraint condition, to obtain the at least one trained first model”. This limitation is an additional element that amounts to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer in its ordinary capacity as a tool to perform an existing process. See MPEP §2106.05(f). Elements that merely amount to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer in its ordinary capacity as a tool to perform an existing process cannot provide an inventive concept. Further, the claim recites: “by the first server”. This limitation is an additional element that amounts to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer in its ordinary capacity as a tool to perform an existing process. See MPEP §2106.05(f). Elements that merely amount to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer in its ordinary capacity as a tool to perform an existing process cannot provide an inventive concept. The claim is not patent eligible.
Regarding claim 5, the rejection of claim 3 is incorporated, and further, the claim recites: “wherein the first distance constraint condition further comprises a constraint formed by a similarity or a distance between the second model and a third model, and the third model is a model from an upstream device of the first server”. This limitation is a continuation of the “training, …, the at least one first model under constraints of a first sparsity constraint condition and a first distance constraint condition, to obtain the at least one trained first model, wherein the first sparsity constraint condition comprises a constraint formed by a function for reducing parameters in the first model, and the first distance constraint condition comprises a constraint formed by a similarity or a distance between the at least one first model and the second model” limitation identified in the rejection of the parent claim. Thus, the claim recites a judicial exception.
The claim does not include any additional elements that amount to an integration of the judicial exception into a practical application, nor to significantly more than the judicial exception. The claim is not patent eligible.
Regarding claim 6, the rejection of claim 1 is incorporated, and further, the claim recites: “receiving, …, a third model from an upstream device”. This limitation is an additional element that amounts to insignificant extra-solution activity to the judicial exception, see MPEP §2106.05(g). Further, this limitation is directed to receiving or transmitting data over a network which courts have recognized as well-understood, routine, and conventional when they are claimed in a generic manner, see MPEP §2106.05(d)(II). Further, the claim recites: “updating, by the first server, a locally stored model to obtain the second model, wherein the updating is based on the third model”. This limitation is an additional element that amounts to is insignificant extra-solution activity to the judicial exception, see MPEP §2106.05(g). Further, this limitation is well‐understood, routine, and conventional as taught by activity is supported under Berkheimer Option 2, Zhan et al., U.S. Patent Application Publication No. 20230325722, Paragraph 0118, Lines 3-17, “A federated learning system usually structurally includes a central server and a plurality of clients as participants. A working procedure mainly includes a model delivery process and a model aggregation process. In the model delivery process, a client downloads a model from the central server, trains, based on locally stored training data, the model downloaded from the central server, and uploads the model to the central server after the model is trained to a specific extent. Training to a specific extent may be understood as a fixed quantity of training rounds. In the model aggregation process, the central server collects models uploaded by clients and aggregates the models. The model delivery process and the model aggregation process are repeated until model convergence”. Further, the claim recites: “by the first server”. This limitation is an additional element that amounts to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer in its ordinary capacity as a tool to perform an existing process. See MPEP §2106.05(f). Elements that merely amount to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer in its ordinary capacity as a tool to perform an existing process cannot provide an inventive concept. The claim is not patent eligible.
Regarding claim 7, the rejection of claim 6 is incorporated, and further, the claim recites: “after the obtaining the second model, the method further comprises: training, …, the third model to obtain an updated third model”. This limitation, under the broadest reasonable interpretation, and according to applicant’s specification, recites mathematical concepts in addition to those identified in the rejection of the parent claim. Thus, the claim recites a judicial exception.
Further, the claim recites: “by the first server”. This limitation is an additional element that amounts to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer in its ordinary capacity as a tool to perform an existing process. See MPEP §2106.05(f). Elements that merely amount to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer in its ordinary capacity as a tool to perform an existing process cannot provide an inventive concept. The claim is not patent eligible.
Regarding claim 8, the rejection of claim 7 is incorporated, and further, the claim recites: “training, …, the third model under a constraint of a second sparsity constraint condition, to obtain the trained third model, wherein the second sparsity constraint condition comprises a constraint formed by a function for reducing parameters in the third model”. This limitation recites mathematical concepts in addition to those identified in the rejection of the parent claim. Thus, the claim recites a judicial exception.
Further, the claim recites: “by the first server”. This limitation is an additional element that amounts to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer in its ordinary capacity as a tool to perform an existing process. See MPEP §2106.05(f). Elements that merely amount to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer in its ordinary capacity as a tool to perform an existing process cannot provide an inventive concept. The claim is not patent eligible.
Regarding claim 9:
Step 1 Statutory Category: Claim 9 is directed to a method, which falls under one of the four statutory categories.
Step 2A Prong 1 Judicial exception: Claim 9 recites, in part, “training, …, the first model under a constraint of a sparsity constraint condition, to obtain a trained first model, wherein the sparsity constraint condition comprises a constraint formed by a function for reducing parameters in the first model”. This limitation, under the broadest reasonable interpretation, covers the recitation of a mathematical concept, see MPEP §2106.04(a)(2)(I).
Step 2A Prong 2 Integration into a practical application: This judicial exception is not integrated into a practical application. In particular the claim recites: “A federated learning method”. This limitation is an additional element that generally links the use of the judicial exception to a particular technological environment or field of use. See MPEP §2106.05(h). Further, the claim recites: “receiving, …, information of a first model, wherein the information of the at least one first model is from a first server”. This limitation is an additional element that amounts to adding insignificant extra-solution activity to the judicial exception. See MPEP §2106.05(g). Further, the claim recites: “sending, …, information of the trained first model to the first server, so that the first server retrains the first model based on the received information of the trained first model, to obtain a new first model”. This limitation is an additional element that amounts to adding insignificant extra-solution activity to the judicial exception. See MPEP §2106.05(g). Further, the claim recites: “by a first client” and “by the first client”. These limitations are additional elements that amount to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer in its ordinary capacity as a tool to perform an existing process. See MPEP §2106.05(f).
Step 2B Significantly more: The claims do not include additional elements that are sufficient to amount to significantly more than the judicial exception. As discussed above with respect to integration of the abstract idea into a practical application, the additional element: “A federated learning method” generally links the use of the judicial exception to a particular technological environment or field of use. Elements that merely generally link the use of the judicial exception to a particular technological environment or field of use cannot provide an inventive concept. Further, the additional elements: “receiving, …, information of a first model, wherein the information of the at least one first model is from a first server” and “sending, …, information of the trained first model to the first server, so that the first server retrains the first model based on the received information of the trained first model, to obtain a new first model” amount to adding insignificant extra-solution activity to the judicial exception, and further, are directed to receiving or transmitting data over a network which courts have recognized as well-understood, routine, and conventional when they are claimed in a generic manner, see MPEP §2106.05(d)(II). Further, the additional elements “by a first client” and “by the first client” amount to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer in its ordinary capacity as a tool to perform an existing process. Elements that merely amount to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer in its ordinary capacity as a tool to perform an existing process cannot provide an inventive concept. The claim is not patent eligible.
Regarding claim 10, the rejection of claim 9 is incorporated, and further, the claim recites: “training, by the first client, the first model under constraints of the sparsity constraint condition and a second distance constraint condition, to obtain the trained first model, wherein the second distance constraint condition comprises a constraint formed by a similarity or a distance between the first model and a locally stored fourth model”. This limitation recites mathematical concepts in addition to those identified in the rejection of the parent claim. Thus, the claim recites a judicial exception.
Further, the claim recites: “by the first client”. This limitation is an additional element that amounts to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer in its ordinary capacity as a tool to perform an existing process. See MPEP §2106.05(f). Elements that merely amount to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer in its ordinary capacity as a tool to perform an existing process cannot provide an inventive concept. The claim is not patent eligible.
Regarding claim 14:
Step 1 Statutory Category: Claim 14 is directed to a machine, which falls under one of the four statutory categories.
Step 2A Prong 1 Judicial exception: Claim 14 recites, in part, “training, …, the at least one first model to obtain at least one trained first model”. This limitation, under the broadest reasonable interpretation, and according to applicant’s specification paragraphs 0008, 0010, 0012, covers the recitation of mathematical concepts, see MPEP §2106.04(a)(2)(I). Further, the claim recites: “aggregating, …, the at least one trained first model to obtain an aggregation result; and updating, …, a locally stored second model to obtain an updated second model, wherein the updating is based on the aggregation result”. This limitation recites, under the broadest reasonable interpretation, covers the recitation of mathematical concepts, see MPEP §2106.04(a)(2)(I).
Step 2A Prong 2 Integration into a practical application: This judicial exception is not integrated into a practical application. In particular the claim recites: “A federated learning apparatus”, “one or more processors”, “a memory coupled to the one or more processors”, “a program”, and “program instructions”. These limitations are additional elements that amount to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer in its ordinary capacity as a tool to perform an existing process. See MPEP §2106.05(f). Further, the claim recites: “receiving, …, information of at least one first model, wherein the information of the at least one first model is from at least one downstream device”. This limitation is an additional element that amounts to adding insignificant extra-solution activity to the judicial exception. See MPEP §2106.05(g). Further, the claim recites: “by a first server” and “by the first server”. These limitations are additional elements that amount to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer in its ordinary capacity as a tool to perform an existing process. See MPEP §2106.05(f).
Step 2B Significantly more: The claims do not include additional elements that are sufficient to amount to significantly more than the judicial exception. As discussed above with respect to integration of the abstract idea into a practical application, the additional elements: “A federated learning apparatus”, “one or more processors”, “a memory coupled to the one or more processors”, “a program”, and “program instructions” amount to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer in its ordinary capacity as a tool to perform an existing process. Elements that merely amount to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer in its ordinary capacity as a tool to perform an existing process cannot provide an inventive concept. Further, the additional element: “receiving, …, information of at least one first model, wherein the information of the at least one first model is from at least one downstream device” is insignificant extra-solution activity to the judicial exception, and further, is directed to receiving or transmitting data over a network which courts have recognized as well-understood, routine, and conventional when they are claimed in a generic manner, see MPEP §2106.05(d)(II). Further, the additional elements “by a first server” and “by the first server” amount to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer in its ordinary capacity as a tool to perform an existing process. Elements that merely amount to adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer in its ordinary capacity as a tool to perform an existing process cannot provide an inventive concept. The claim is not patent eligible.
Regarding claim 15, the rejection of claim 14 is incorporated, and further, claim 15 is substantially similar to claim 2 respectively, and is rejected in the same manner and reasoning applying.
Regarding claim 16, the rejection of claim 14 is incorporated, and further, claim 16 is substantially similar to claim 3 respectively, and is rejected in the same manner and reasoning applying.
Regarding claim 17, the rejection of claim 16 is incorporated, and further, claim 17 is substantially similar to claim 4 respectively, and is rejected in the same manner and reasoning applying.
Regarding claim 18, the rejection of claim 16 is incorporated, and further, claim 18 is substantially similar to claim 5 respectively, and is rejected in the same manner and reasoning applying.
Regarding claim 19, the rejection of claim 14 is incorporated, and further, claim 19 is substantially similar to claim 6 respectively, and is rejected in the same manner and reasoning applying.
Regarding claim 20, the rejection of claim 19 is incorporated, and further, claim 20 is substantially similar to claim 7 respectively, and is rejected in the same manner and reasoning applying.
Claim Rejections - 35 USC § 102
The following is a quotation of the appropriate paragraphs of 35 U.S.C. 102 that form the basis for the rejections under this section made in this Office action:
A person shall be entitled to a patent unless –
(a)(1) the claimed invention was patented, described in a printed publication, or in public use, on sale, or otherwise available to the public before the effective filing date of the claimed invention.
Claims 1, 6-7, 9, 11, 14, and 19-20 is/are rejected under 35 U.S.C. 102(a)(1) as being anticipated by Thapa et al., SplitFed: When Federated Learning Meets Split Learning, 09/02/2020, https://arxiv.org/pdf/2004.12088v2, hereinafter referred to as “Thapa”.
Regarding claim 1, Thapa teaches A federated learning method (Thapa, Page 1, Abstract, Lines 8-9, “we present a novel approach, named splitfed (SFL), that amalgamates the two approaches eliminating their inherent drawbacks”), comprising:
receiving, by a first server, information of at least one first model, wherein the information of the at least one first model is from at least one downstream device (Thapa, Page 5, Section 3.1, Lines 4-6, “all clients (e.g., Hospitals, IoMTs with low computing resources) carry out the forward propagations on their client-side model in parallel, then pass their smashed data to the (main) server”);
training, by the first server, the at least one first model to obtain at least one trained first model (Thapa, Pages 5-6, Section 3.1, Lines 6-9, “Then the server, which is assumed to have sufficient computing resources (e.g., cloud server and researchers with high-performance computing resources), process the forward propagation and back-propagation on its server-side model with each client's smashed data in (somewhat) parallel”);
aggregating, by the first server, the at least one trained first model to obtain an aggregation result; and updating, by the first server, a locally stored second model to obtain an updated second model, wherein the updating is based on the aggregation result (Thapa, Page 6, Lines 3-4, “Afterward, the server updates its model by a weighted average of the gradients that it computes during the back-propagation on each client's smashed data”).
Regarding claim 6, the rejection of claim 1 is incorporated, and further, Thapa teaches wherein the method further comprises:
receiving, by the first server, a third model from an upstream device (Thapa, Page 5, Section 3.1, Lines 4-6, “all clients (e.g., Hospitals, IoMTs with low computing resources) carry out the forward propagations on their client-side model in parallel, then pass their smashed data to the (main) server”; see also Thapa, Page 5, Figure 2, Client 1 is considered to be “upstream” while client k is considered to be “downstream”); and
updating, by the first server, a locally stored model to obtain the second model, wherein the updating is based on the third model (Thapa, Page 6, Lines 3-4, “Afterward, the server updates its model by a weighted average of the gradients that it computes during the back-propagation on each client's smashed data”).
Regarding claim 7, the rejection of claim 6 is incorporated, and further, Thapa teaches after the obtaining the second model, the method further comprises: training, by the first server, the third model to obtain an updated third model (Thapa, Pages 5-6, Section 3.1, Lines 6-9, “Then the server, which is assumed to have sufficient computing resources (e.g., cloud server and researchers with high-performance computing resources), process the forward propagation and back-propagation on its server-side model with each client's smashed data in (somewhat) parallel”).
Regarding claim 9, Thapa teaches A federated learning method, comprising: receiving, by a first client, information of a first model, wherein the information of the at least one first model is from a first server (Thapa, Page 6, Algorithm 3, “Lines 1-4, “/*Runs on Main Server */ / EnsureMainServer executes at round t
≥
0: / for each client
k
∈
S
t
in parallel do / (
A
k
,
t
,
Y
k
)
←
C
l
i
e
n
t
U
p
d
a
t
e
(
W
k
,
t
C
)
”);
training, by the first client, the first model under a constraint of a sparsity constraint condition, to obtain a trained first model, wherein the sparsity constraint condition comprises a constraint formed by a function for reducing parameters in the first model (Thapa, Page 8, Section 5.2, Paragraph 2, Lines 2-5, and Equations 1 and 2, “Now, firstly, after t time, the client k, receives the gradients
d
A
k
,
t
from the server, and with this, it calculates client-side gradients
∇
l
k
(
W
k
,
i
,
t
C
)
for each of its local sample
x
i
, and
g
k
,
t
x
i
←
∇
l
k
W
k
,
i
,
t
C
. Secondly, the
l
2
-norm of each gradient is clipped according to the following equation:
g
-
k
,
t
(
x
i
)
←
g
k
,
t
x
i
/
max
1
,
g
k
,
t
x
i
2
C
'
”); and
sending, by the first client, information of the trained first model to the first server, so that the first server retrains the first model based on the received information of the trained first model, to obtain a new first model (Thapa, Page 5, Section 3.1, Lines 4-6, “all clients (e.g., Hospitals, IoMTs with low computing resources) carry out the forward propagations on their client-side model in parallel, then pass their smashed data to the (main) server”; Thapa, Pages 5-6, Section 3.1, Lines 6-9, “Then the server, which is assumed to have sufficient computing resources (e.g., cloud server and researchers with high-performance computing resources), process the forward propagation and back-propagation on its server-side model with each client's smashed data in (somewhat) parallel”).
Regarding claim 11, Thapa teaches A federated learning system, comprising a plurality of servers and at least one client (Thapa, Page 1, Abstract, Lines 8-9, “we present a novel approach, named splitfed (SFL), that amalgamates the two approaches eliminating their inherent drawbacks”; Thapa, Page 5, Figure 2, a plurality of servers and a plurality of clients can be seen), wherein
any one of the plurality of servers is configured to implement a federated learning method for a first server, and the at least one client is configured to implement a federated learning method for a first client(Thapa, Page 1, Abstract, Lines 8-9, “we present a novel approach, named splitfed (SFL), that amalgamates the two approaches eliminating their inherent drawbacks”; Thapa, Page 5, Figure 2, a plurality of servers and a plurality of clients can be seen; see also Thapa, Page 6, Algorithm 3), wherein the federated learning method for the first server comprises:
receiving, by the first server, information of at least one first model, wherein the information of the at least one first model is from at least one downstream device (Thapa, Page 5, Section 3.1, Lines 4-6, “all clients (e.g., Hospitals, IoMTs with low computing resources) carry out the forward propagations on their client-side model in parallel, then pass their smashed data to the (main) server”);
training, by the first server, the at least one first model to obtain at least one trained first model (Thapa, Pages 5-6, Section 3.1, Lines 6-9, “Then the server, which is assumed to have sufficient computing resources (e.g., cloud server and researchers with high-performance computing resources), process the forward propagation and back-propagation on its server-side model with each client's smashed data in (somewhat) parallel”);
aggregating, by the first server, the at least one trained first model to obtain an aggregation result; and updating, by the first server, a locally stored second model to obtain an updated second model, wherein the updating is based on the aggregation result (Thapa, Page 6, Lines 3-4, “Afterward, the server updates its model by a weighted average of the gradients that it computes during the back-propagation on each client's smashed data”);
and the federated learning method for the first client comprises:
receiving, by the first client, information of the updated second model, wherein the information of the updated second model is from the first server (Thapa, Page 6, Algorithm 3, “Lines 1-4, “/*Runs on Main Server */ / EnsureMainServer executes at round t
≥
0: / for each client
k
∈
S
t
in parallel do / (
A
k
,
t
,
Y
k
)
←
C
l
i
e
n
t
U
p
d
a
t
e
(
W
k
,
t
C
)
”);
training, by the first client, the updated second model under a constraint of a first sparsity constraint condition, to obtain a trained updated second model, wherein the first sparsity constraint condition comprises a constraint formed by a function for reducing parameters in the updated second model (Thapa, Page 8, Section 5.2, Paragraph 2, Lines 2-5, and Equations 1 and 2, “Now, firstly, after t time, the client k, receives the gradients
d
A
k
,
t
from the server, and with this, it calculates client-side gradients
∇
l
k
(
W
k
,
i
,
t
C
)
for each of its local sample
x
i
, and
g
k
,
t
x
i
←
∇
l
k
W
k
,
i
,
t
C
. Secondly, the
l
2
-norm of each gradient is clipped according to the following equation:
g
-
k
,
t
(
x
i
)
←
g
k
,
t
x
i
/
max
1
,
g
k
,
t
x
i
2
C
'
”); and
sending, by the first client, information about the trained updated second model to the first server, so that the first server retrains the updated second model based on the received information of the trained updated second model, to obtain a new updated second model (Thapa, Page 5, Section 3.1, Lines 4-6, “all clients (e.g., Hospitals, IoMTs with low computing resources) carry out the forward propagations on their client-side model in parallel, then pass their smashed data to the (main) server”; Thapa, Pages 5-6, Section 3.1, Lines 6-9, “Then the server, which is assumed to have sufficient computing resources (e.g., cloud server and researchers with high-performance computing resources), process the forward propagation and back-propagation on its server-side model with each client's smashed data in (somewhat) parallel”).
Regarding claim 14, Thapa teaches A federated learning apparatus, comprising one or more processors and a memory coupled to the one or more processors, the memory stores a program that, when program instructions of the program are executed by the one or more processors (Thapa, Page 9, Section 6, Lines 1-7, “Experiments are carried out on uniformly distributed and horizontally partitioned image datasets among clients. All programs are written in python 3.7.2 using the PyTorch library (PyTorch 1.2.0). For quicker experiments and developments, we use the High-Performance Computing (HPC) platform that is built on Dell EMC's PowerEdge platform with partner GPUs for computation and InfiniBand networking. We run clients and servers on different computing nodes of the cluster provided by HPC. We request the following resources for one slurm job on HPC: 10GB of RAM, one GPU (Tesla P100-SXM2-16GB), one computing node with at least one task per node. The architecture of the nodes is x86_64”, cause the federated learning apparatus to perform a method (Thapa, Page 1, Abstract, Lines 8-9, “we present a novel approach, named splitfed (SFL), that amalgamates the two approaches eliminating their inherent drawbacks”) comprising:
receiving, by a first server, information of at least one first model, wherein the information of the at least one first model is from at least one downstream device (Thapa, Page 5, Section 3.1, Lines 4-6, “all clients (e.g., Hospitals, IoMTs with low computing resources) carry out the forward propagations on their client-side model in parallel, then pass their smashed data to the (main) server”);
training, by the first server, the at least one first model to obtain at least one trained first model (Thapa, Pages 5-6, Section 3.1, Lines 6-9, “Then the server, which is assumed to have sufficient computing resources (e.g., cloud server and researchers with high-performance computing resources), process the forward propagation and back-propagation on its server-side model with each client's smashed data in (somewhat) parallel”);
aggregating, by the first server, the at least one trained first model to obtain an aggregation result; and updating, by the first server, a locally stored second model to obtain an updated second model, wherein the updating is based on the aggregation result (Thapa, Page 6, Lines 3-4, “Afterward, the server updates its model by a weighted average of the gradients that it computes during the back-propagation on each client's smashed data”).
Regarding claim 19, the rejection of claim 14 is incorporated, and further, Thapa teaches wherein the method further comprises:
receiving, by the first server, a third model from an upstream device (Thapa, Page 5, Section 3.1, Lines 4-6, “all clients (e.g., Hospitals, IoMTs with low computing resources) carry out the forward propagations on their client-side model in parallel, then pass their smashed data to the (main) server”; see also Thapa, Page 5, Figure 2, Client 1 is considered to be “upstream” while client k is considered to be “downstream”); and
updating, by the first server, a locally stored model to obtain the second model, wherein the updating is based on the third model (Thapa, Page 6, Lines 3-4, “Afterward, the server updates its model by a weighted average of the gradients that it computes during the back-propagation on each client's smashed data”).
Regarding claim 20, the rejection of claim 19 is incorporated, and further, Thapa teaches after the obtaining the second model, the method further comprises: training, by the first server, the third model to obtain an updated third model (Thapa, Pages 5-6, Section 3.1, Lines 6-9, “Then the server, which is assumed to have sufficient computing resources (e.g., cloud server and researchers with high-performance computing resources), process the forward propagation and back-propagation on its server-side model with each client's smashed data in (somewhat) parallel”).
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 2, 8, 12, and 15 are rejected under 35 U.S.C. 103 as being unpatentable over Thapa in view of Naseri et al., Toward Robustness and Privacy in Federated Learning: Experimenting with Local and Central Differential Privacy, 03/15/2021, https://arxiv.org/pdf/2009.03561v3, hereinafter referred to as “Naseri”.
Regarding claim 2, the rejection of claim 1 is incorporated, and further, Thapa teaches training, …, the at least one first model under a constraint of a first sparsity constraint condition, to obtain the at least one trained first model, wherein the first sparsity constraint condition comprises a constraint formed by a function for reducing parameters in the first model (Thapa, Page 8, Section 5.2, Paragraph 2, Lines 2-5, and Equations 1 and 2, “Now, firstly, after t time, the client k, receives the gradients
d
A
k
,
t
from the server, and with this, it calculates client-side gradients
∇
l
k
(
W
k
,
i
,
t
C
)
for each of its local sample
x
i
, and
g
k
,
t
x
i
←
∇
l
k
W
k
,
i
,
t
C
. Secondly, the
l
2
-norm of each gradient is clipped according to the following equation:
g
-
k
,
t
(
x
i
)
←
g
k
,
t
x
i
/
max
1
,
g
k
,
t
x
i
2
C
'
”).
Thapa does not explicitly teach that this training is done by the first server as part of the server-side model training.
Naseri teaches performing the differential privacy training by the first server (Naseri, Pages 3-4, Section 2.3, Central Differential Privacy (CDP), Paragraph 2, Lines 1-5, “In this paper, we implement the CDP approach for FL dis cussed in [68] and [33], which is illustrated in Algorithm 2. The server clips the l2 norm of participants’ updates, then, it aggregates the clipped updates and adds Gaussian noise to the aggregate”).
It would have been obvious to a person of ordinary skill in the art, before the effective filing date of the invention, to have modified the federated learning method of Thapa to include performing central differential privacy during training at the server as taught by Naseri. The motivation to do so would have been to prevent the server model from overfitting to any client’s update and central differential privacy yields better utility (Naseri, Page 4, Lines 2-3, “This prevents overfitting to any participant’s updates”; Naseri, Page 9, Section 3.6, Paragraph 3, Lines 1-2, “Overall, CDP works better as it better mitigates the attack and yields better utility”).
Regarding claim 8, the rejection of claim 7 is incorporated, and further, Thapa teaches training, …, the third model under a constraint of a second sparsity constraint condition, to obtain the trained third model, wherein the second sparsity constraint condition comprises a constraint formed by a function for reducing parameters in the third model (Thapa, Page 8, Section 5.2, Paragraph 2, Lines 2-5, and Equations 1 and 2, “Now, firstly, after t time, the client k, receives the gradients
d
A
k
,
t
from the server, and with this, it calculates client-side gradients
∇
l
k
(
W
k
,
i
,
t
C
)
for each of its local sample
x
i
, and
g
k
,
t
x
i
←
∇
l
k
W
k
,
i
,
t
C
. Secondly, the
l
2
-norm of each gradient is clipped according to the following equation:
g
-
k
,
t
(
x
i
)
←
g
k
,
t
x
i
/
max
1
,
g
k
,
t
x
i
2
C
'
”).
Thapa does not explicitly teach that this training is done by the first server as part of the server-side model training.
Naseri teaches performing the differential privacy training by the first server (Naseri, Pages 3-4, Section 2.3, Central Differential Privacy (CDP), Paragraph 2, Lines 1-5, “In this paper, we implement the CDP approach for FL dis cussed in [68] and [33], which is illustrated in Algorithm 2. The server clips the l2 norm of participants’ updates, then, it aggregates the clipped updates and adds Gaussian noise to the aggregate”).
It would have been obvious to a person of ordinary skill in the art, before the effective filing date of the invention, to have modified the federated learning method of Thapa to include performing central differential privacy during training at the server as taught by Naseri. The motivation to do so would have been to prevent the server model from overfitting to any client’s update and central differential privacy yields better utility (Naseri, Page 4, Lines 2-3, “This prevents overfitting to any participant’s updates”; Naseri, Page 9, Section 3.6, Paragraph 3, Lines 1-2, “Overall, CDP works better as it better mitigates the attack and yields better utility”).
Regarding claim 12, the rejection of claim 11 is incorporated, and further, Thapa teaches training, …, the at least one first model under constraints of a second sparsity constraint condition, to obtain the at least one trained first model, wherein the second sparsity constraint condition comprises a constraint formed by a function for reducing parameters in the first model (Thapa, Page 8, Section 5.2, Paragraph 2, Lines 2-5, and Equations 1 and 2, “Now, firstly, after t time, the client k, receives the gradients
d
A
k
,
t
from the server, and with this, it calculates client-side gradients
∇
l
k
(
W
k
,
i
,
t
C
)
for each of its local sample
x
i
, and
g
k
,
t
x
i
←
∇
l
k
W
k
,
i
,
t
C
. Secondly, the
l
2
-norm of each gradient is clipped according to the following equation:
g
-
k
,
t
(
x
i
)
←
g
k
,
t
x
i
/
max
1
,
g
k
,
t
x
i
2
C
'
”).
Thapa does not explicitly teach that this training is done by the first server as part of the server-side model training.
Naseri teaches performing the differential privacy training by the first server (Naseri, Pages 3-4, Section 2.3, Central Differential Privacy (CDP), Paragraph 2, Lines 1-5, “In this paper, we implement the CDP approach for FL dis cussed in [68] and [33], which is illustrated in Algorithm 2. The server clips the l2 norm of participants’ updates, then, it aggregates the clipped updates and adds Gaussian noise to the aggregate”).
It would have been obvious to a person of ordinary skill in the art, before the effective filing date of the invention, to have modified the federated learning method of Thapa to include performing central differential privacy during training at the server as taught by Naseri. The motivation to do so would have been to prevent the server model from overfitting to any client’s update and central differential privacy yields better utility (Naseri, Page 4, Lines 2-3, “This prevents overfitting to any participant’s updates”; Naseri, Page 9, Section 3.6, Paragraph 3, Lines 1-2, “Overall, CDP works better as it better mitigates the attack and yields better utility”).
It is noted the claim recites alternative language and the proposed combination teaches at least one of the recited alternatives.
Regarding claim 15, the rejection of claim 14 is incorporated, and further, Thapa teaches training, …, the at least one first model under a constraint of a first sparsity constraint condition, to obtain the at least one trained first model, wherein the first sparsity constraint condition comprises a constraint formed by a function for reducing parameters in the first model (Thapa, Page 8, Section 5.2, Paragraph 2, Lines 2-5, and Equations 1 and 2, “Now, firstly, after t time, the client k, receives the gradients
d
A
k
,
t
from the server, and with this, it calculates client-side gradients
∇
l
k
(
W
k
,
i
,
t
C
)
for each of its local sample
x
i
, and
g
k
,
t
x
i
←
∇
l
k
W
k
,
i
,
t
C
. Secondly, the
l
2
-norm of each gradient is clipped according to the following equation:
g
-
k
,
t
(
x
i
)
←
g
k
,
t
x
i
/
max
1
,
g
k
,
t
x
i
2
C
'
”).
Thapa does not explicitly teach that this training is done by the first server as part of the server-side model training.
Naseri teaches performing the differential privacy training by the first server (Naseri, Pages 3-4, Section 2.3, Central Differential Privacy (CDP), Paragraph 2, Lines 1-5, “In this paper, we implement the CDP approach for FL dis cussed in [68] and [33], which is illustrated in Algorithm 2. The server clips the l2 norm of participants’ updates, then, it aggregates the clipped updates and adds Gaussian noise to the aggregate”).
It would have been obvious to a person of ordinary skill in the art, before the effective filing date of the invention, to have modified the federated learning method of Thapa to include performing central differential privacy during training at the server as taught by Naseri. The motivation to do so would have been to prevent the server model from overfitting to any client’s update and central differential privacy yields better utility (Naseri, Page 4, Lines 2-3, “This prevents overfitting to any participant’s updates”; Naseri, Page 9, Section 3.6, Paragraph 3, Lines 1-2, “Overall, CDP works better as it better mitigates the attack and yields better utility”).
Claim 10 is rejected under 35 U.S.C. 103 as being unpatentable over Thapa in view of He et al., Group Knowledge Transfer: Federated Learning of Large CNNs at the Edge, 11/05/2020, https://arxiv.org/pdf/2007.14513, hereinafter referred to as “He”.
Regarding claim 10, the rejection of claim 9 is incorporated, and further, Thapa teaches training, by the first client, the first model under constraints of the sparsity constraint condition …, to obtain the trained first model (Thapa, Page 8, Section 5.2, Paragraph 2, Lines 2-5, and Equations 1 and 2, “Now, firstly, after t time, the client k, receives the gradients
d
A
k
,
t
from the server, and with this, it calculates client-side gradients
∇
l
k
(
W
k
,
i
,
t
C
)
for each of its local sample
x
i
, and
g
k
,
t
x
i
←
∇
l
k
W
k
,
i
,
t
C
. Secondly, the
l
2
-norm of each gradient is clipped according to the following equation:
g
-
k
,
t
(
x
i
)
←
g
k
,
t
x
i
/
max
1
,
g
k
,
t
x
i
2
C
'
”).
Thapa does not explicitly teach a second distance constraint condition … wherein the second distance constraint condition comprises a constraint formed by a similarity or a distance between the first model and a locally stored fourth model.
He teaches a second distance constraint condition … wherein the second distance constraint condition comprises a constraint formed by a similarity between the first model and a locally stored fourth model (He, Page 5, Lines 3-5 and Equation 7, “The server CNN absorbs the knowledge from many edges, and an individual edge CNN obtains enhanced knowledge from the server CNN. To be more specific, in Eq. (2) and (5), we design…
l
c
as follows.
l
c
(
k
)
=
l
C
E
+
l
K
D
z
s
,
z
c
k
=
l
C
E
+
D
K
L
(
p
k
|
|
p
s
)
”; He, Page 5, Lines 6-7, “
D
K
L
is the Kullback Leibler (KL) Divergence function that serves as a term in the loss function
l
s
and
l
c
to transfer knowledge from a network to another”).
It would have been obvious to a person of ordinary skill in the art, before the effective filing date of the invention, to have modified the federated learning method of Thapa to include training the model under a distance constraint condition comprising a constraint formed by a similarity as taught by He. The motivation to do so would have been to improve the models’ feature extraction capability and boost performance of the client and server models (He, Page 5, Paragraph 2, “Intuitively, the KL divergence loss attempts to bring the soft label and the ground truth close to each other. In doing so, the server model absorbs the knowledge gained from each of the edge models. Similarly, the edge models attempt to bring their predictions closer to the server model’s prediction and thereby absorb the server model knowledge to improve their feature extraction capability”; He, Page 2, Paragraph 2, Lines 8-10, “Under this reformulation, FedGKT not only boosts training CNNs at the edge but also contributes to the development of a new knowledge distillation (KD) paradigm, group knowledge transfer, to boost the performance of the server model”).
It is noted the claim recites alternative language and the proposed combination teaches at least one of the recited alternatives.
Claims 3-5, 13, and 16-18 are rejected under 35 U.S.C. 103 as being unpatentable over Thapa in view of Naseri in further view of He.
Regarding claim 3, the rejection of claim 1 is incorporated, and further, Thapa teaches training, …, the at least one first model under constraints of a first sparsity constraint condition …, to obtain the at least one trained first model, wherein the first sparsity constraint condition comprises a constraint formed by a function for reducing parameters in the first model (Thapa, Page 8, Section 5.2, Paragraph 2, Lines 2-5, and Equations 1 and 2, “Now, firstly, after t time, the client k, receives the gradients
d
A
k
,
t
from the server, and with this, it calculates client-side gradients
∇
l
k
(
W
k
,
i
,
t
C
)
for each of its local sample
x
i
, and
g
k
,
t
x
i
←
∇
l
k
W
k
,
i
,
t
C
. Secondly, the
l
2
-norm of each gradient is clipped according to the following equation:
g
-
k
,
t
(
x
i
)
←
g
k
,
t
x
i
/
max
1
,
g
k
,
t
x
i
2
C
'
”).
Thapa does not explicitly teach that this training is done by the first server as part of the server-side model training nor a first distance constraint condition … and the first distance constraint condition comprises a constraint formed by a similarity or a distance between the at least one first model and the second model.
Naseri teaches performing the differential privacy training by the first server (Naseri, Pages 3-4, Section 2.3, Central Differential Privacy (CDP), Paragraph 2, Lines 1-5, “In this paper, we implement the CDP approach for FL dis cussed in [68] and [33], which is illustrated in Algorithm 2. The server clips the l2 norm of participants’ updates, then, it aggregates the clipped updates and adds Gaussian noise to the aggregate”).
It would have been obvious to a person of ordinary skill in the art, before the effective filing date of the invention, to have modified the federated learning method of Thapa to include performing central differential privacy during training at the server as taught by Naseri. The motivation to do so would have been to prevent the server model from overfitting to any client’s update and central differential privacy yields better utility (Naseri, Page 4, Lines 2-3, “This prevents overfitting to any participant’s updates”; Naseri, Page 9, Section 3.6, Paragraph 3, Lines 1-2, “Overall, CDP works better as it better mitigates the attack and yields better utility”).
Thapa in view of He does not explicitly teach a first distance constraint condition … and the first distance constraint condition comprises a constraint formed by a similarity or a distance between the at least one first model and the second model.
He teaches a first distance constraint condition … and the first distance constraint condition comprises a constraint formed by a similarity between the at least one first model and the second model (He, Page 5, Lines 3-5 and Equation 7, “The server CNN absorbs the knowledge from many edges, and an individual edge CNN obtains enhanced knowledge from the server CNN. To be more specific, in Eq. (2) and (5), we design…
l
c
as follows.
l
c
(
k
)
=
l
C
E
+
l
K
D
z
s
,
z
c
k
=
l
C
E
+
D
K
L
(
p
k
|
|
p
s
)
”; He, Page 5, Lines 6-7, “
D
K
L
is the Kullback Leibler (KL) Divergence function that serves as a term in the loss function
l
s
and
l
c
to transfer knowledge from a network to another”).
It would have been obvious to a person of ordinary skill in the art, before the effective filing date of the invention, to have modified the federated learning method of the proposed combination to include training the model under a distance constraint condition comprising a constraint formed by a similarity as taught by He. The motivation to do so would have been to improve the models’ feature extraction capability and boost performance of the client and server models (He, Page 5, Paragraph 2, “Intuitively, the KL divergence loss attempts to bring the soft label and the ground truth close to each other. In doing so, the server model absorbs the knowledge gained from each of the edge models. Similarly, the edge models attempt to bring their predictions closer to the server model’s prediction and thereby absorb the server model knowledge to improve their feature extraction capability”; He, Page 2, Paragraph 2, Lines 8-10, “Under this reformulation, FedGKT not only boosts training CNNs at the edge but also contributes to the development of a new knowledge distillation (KD) paradigm, group knowledge transfer, to boost the performance of the server model”).
It is noted the claim recites alternative language and the proposed combination teaches at least one of the recited alternatives.
Regarding claim 4, the rejection of claim 3 is incorporated, and further, the proposed combination teaches perturbing, by the first server (Naseri, Pages 3-4, Section 2.3, Central Differential Privacy (CDP), Paragraph 2, Lines 1-5, “In this paper, we implement the CDP approach for FL dis cussed in [68] and [33], which is illustrated in Algorithm 2. The server clips the l2 norm of participants’ updates, then, it aggregates the clipped updates and adds Gaussian noise to the aggregate”), the at least one first model by using noise, to obtain at least one updated first model (Thapa, Page 8, Section 5.2, Paragraph 2, Line 6 and Equation 3, “Thirdly, calibrated noise is added to the average gradient:
g
~
k
,
t
←
1
n
k
∑
i
g
-
k
,
t
x
i
+
N
0
,
σ
2
C
'
2
I
”); and
training, by the first server (Naseri, Pages 3-4, Section 2.3, Central Differential Privacy (CDP), Paragraph 2, Lines 1-5, “In this paper, we implement the CDP approach for FL dis cussed in [68] and [33], which is illustrated in Algorithm 2. The server clips the l2 norm of participants’ updates, then, it aggregates the clipped updates and adds Gaussian noise to the aggregate”), the at least one updated first model under the constraints of the first sparsity constraint condition and the first distance constraint condition, to obtain the at least one trained first model (Thapa, Page 8, Section 5.2, Paragraph 2, Lines 2-5, and Equations 1 and 2, “Now, firstly, after t time, the client k, receives the gradients
d
A
k
,
t
from the server, and with this, it calculates client-side gradients
∇
l
k
(
W
k
,
i
,
t
C
)
for each of its local sample
x
i
, and
g
k
,
t
x
i
←
∇
l
k
W
k
,
i
,
t
C
. Secondly, the
l
2
-norm of each gradient is clipped according to the following equation:
g
-
k
,
t
(
x
i
)
←
g
k
,
t
x
i
/
max
1
,
g
k
,
t
x
i
2
C
'
”; He, Page 5, Lines 3-5 and Equation 6, “The server CNN absorbs the knowledge from many edges…To be more specific, in Eq. (2) and (5), we design
l
s
… as follows.
l
s
=
l
C
E
+
∑
k
=
1
K
l
K
D
z
s
,
z
c
k
=
l
C
E
+
∑
k
=
1
K
D
K
L
(
p
k
|
|
p
s
)
”; He, Page 5, Lines 6-7, “
D
K
L
is the Kullback Leibler (KL) Divergence function that serves as a term in the loss function
l
s
and
l
c
to transfer knowledge from a network to another”).
Regarding claim 5, the rejection of claim 3 is incorporated, and further, the proposed combination teaches wherein the first distance constraint condition further comprises a constraint formed by a similarity between the second model and a third model, and the third model is a model from an upstream device of the first server (Thapa, Page 6, Lines 3-4, “Afterward, the server updates its model by a weighted average of the gradients that it computes during the back-propagation on each client's smashed data”; see also Thapa, Page 5, Figure 2, Client 1 is considered to be “upstream” while client k is considered to be “downstream”; He, Page 5, Lines 3-5 and Equation 6, “The server CNN absorbs the knowledge from many edges…To be more specific, in Eq. (2) and (5), we design
l
s
… as follows.
l
s
=
l
C
E
+
∑
k
=
1
K
l
K
D
z
s
,
z
c
k
=
l
C
E
+
∑
k
=
1
K
D
K
L
(
p
k
|
|
p
s
)
”; He, Page 5, Lines 6-7, “
D
K
L
is the Kullback Leibler (KL) Divergence function that serves as a term in the loss function
l
s
and
l
c
to transfer knowledge from a network to another”; The “server CNN absorbs the knowledge from many edges”, thus, in combination with Thapa, the server CNN would absorb knowledge from upstream and downstream devices).
It is noted the claim recites alternative language and the proposed combination teaches at least one of the recited alternatives.
Regarding claim 13, the rejection of claim 12 is incorporated, and further, the proposed combination teaches perturbing, by the first server (Naseri, Pages 3-4, Section 2.3, Central Differential Privacy (CDP), Paragraph 2, Lines 1-5, “In this paper, we implement the CDP approach for FL dis cussed in [68] and [33], which is illustrated in Algorithm 2. The server clips the l2 norm of participants’ updates, then, it aggregates the clipped updates and adds Gaussian noise to the aggregate”), the at least one first model by using noise, to obtain at least one updated first model (Thapa, Page 8, Section 5.2, Paragraph 2, Line 6 and Equation 3, “Thirdly, calibrated noise is added to the average gradient:
g
~
k
,
t
←
1
n
k
∑
i
g
-
k
,
t
x
i
+
N
0
,
σ
2
C
'
2
I
”); and
training, by the first server (Naseri, Pages 3-4, Section 2.3, Central Differential Privacy (CDP), Paragraph 2, Lines 1-5, “In this paper, we implement the CDP approach for FL dis cussed in [68] and [33], which is illustrated in Algorithm 2. The server clips the l2 norm of participants’ updates, then, it aggregates the clipped updates and adds Gaussian noise to the aggregate”), the at least one updated first model under the constraints of the second sparsity constraint condition …, to obtain the at least one trained first model (Thapa, Page 8, Section 5.2, Paragraph 2, Lines 2-5, and Equations 1 and 2, “Now, firstly, after t time, the client k, receives the gradients
d
A
k
,
t
from the server, and with this, it calculates client-side gradients
∇
l
k
(
W
k
,
i
,
t
C
)
for each of its local sample
x
i
, and
g
k
,
t
x
i
←
∇
l
k
W
k
,
i
,
t
C
. Secondly, the
l
2
-norm of each gradient is clipped according to the following equation:
g
-
k
,
t
(
x
i
)
←
g
k
,
t
x
i
/
max
1
,
g
k
,
t
x
i
2
C
'
”)
The proposed combination thus far does not explicitly teach a first distance constraint condition.
He teaches a first distance constraint condition (He, Page 5, Lines 3-5 and Equation 6, “The server CNN absorbs the knowledge from many edges…To be more specific, in Eq. (2) and (5), we design
l
s
… as follows.
l
s
=
l
C
E
+
∑
k
=
1
K
l
K
D
z
s
,
z
c
k
=
l
C
E
+
∑
k
=
1
K
D
K
L
(
p
k
|
|
p
s
)
”; He, Page 5, Lines 6-7, “
D
K
L
is the Kullback Leibler (KL) Divergence function that serves as a term in the loss function
l
s
and
l
c
to transfer knowledge from a network to another”).
It would have been obvious to a person of ordinary skill in the art, before the effective filing date of the invention, to have modified the federated learning method of the proposed combination to include training the model under a distance constraint condition comprising a constraint formed by a similarity as taught by He. The motivation to do so would have been to improve the models’ feature extraction capability and boost performance of the client and server models (He, Page 5, Paragraph 2, “Intuitively, the KL divergence loss attempts to bring the soft label and the ground truth close to each other. In doing so, the server model absorbs the knowledge gained from each of the edge models. Similarly, the edge models attempt to bring their predictions closer to the server model’s prediction and thereby absorb the server model knowledge to improve their feature extraction capability”; He, Page 2, Paragraph 2, Lines 8-10, “Under this reformulation, FedGKT not only boosts training CNNs at the edge but also contributes to the development of a new knowledge distillation (KD) paradigm, group knowledge transfer, to boost the performance of the server model”).
Regarding claim 16, the rejection of claim 14 is incorporated, and further, Thapa teaches training, …, the at least one first model under constraints of a first sparsity constraint condition …, to obtain the at least one trained first model, wherein the first sparsity constraint condition comprises a constraint formed by a function for reducing parameters in the first model (Thapa, Page 8, Section 5.2, Paragraph 2, Lines 2-5, and Equations 1 and 2, “Now, firstly, after t time, the client k, receives the gradients
d
A
k
,
t
from the server, and with this, it calculates client-side gradients
∇
l
k
(
W
k
,
i
,
t
C
)
for each of its local sample
x
i
, and
g
k
,
t
x
i
←
∇
l
k
W
k
,
i
,
t
C
. Secondly, the
l
2
-norm of each gradient is clipped according to the following equation:
g
-
k
,
t
(
x
i
)
←
g
k
,
t
x
i
/
max
1
,
g
k
,
t
x
i
2
C
'
”).
Thapa does not explicitly teach that this training is done by the first server as part of the server-side model training nor a first distance constraint condition … and the first distance constraint condition comprises a constraint formed by a similarity or a distance between the at least one first model and the second model.
Naseri teaches performing the differential privacy training by the first server (Naseri, Pages 3-4, Section 2.3, Central Differential Privacy (CDP), Paragraph 2, Lines 1-5, “In this paper, we implement the CDP approach for FL dis cussed in [68] and [33], which is illustrated in Algorithm 2. The server clips the l2 norm of participants’ updates, then, it aggregates the clipped updates and adds Gaussian noise to the aggregate”).
It would have been obvious to a person of ordinary skill in the art, before the effective filing date of the invention, to have modified the federated learning method of Thapa to include performing central differential privacy during training at the server as taught by Naseri. The motivation to do so would have been to prevent the server model from overfitting to any client’s update and central differential privacy yields better utility (Naseri, Page 4, Lines 2-3, “This prevents overfitting to any participant’s updates”; Naseri, Page 9, Section 3.6, Paragraph 3, Lines 1-2, “Overall, CDP works better as it better mitigates the attack and yields better utility”).
Thapa in view of He does not explicitly teach a first distance constraint condition … and the first distance constraint condition comprises a constraint formed by a similarity or a distance between the at least one first model and the second model.
He teaches a first distance constraint condition … and the first distance constraint condition comprises a constraint formed by a similarity between the at least one first model and the second model (He, Page 5, Lines 3-5 and Equation 7, “The server CNN absorbs the knowledge from many edges, and an individual edge CNN obtains enhanced knowledge from the server CNN. To be more specific, in Eq. (2) and (5), we design…
l
c
as follows.
l
c
(
k
)
=
l
C
E
+
l
K
D
z
s
,
z
c
k
=
l
C
E
+
D
K
L
(
p
k
|
|
p
s
)
”; He, Page 5, Lines 6-7, “
D
K
L
is the Kullback Leibler (KL) Divergence function that serves as a term in the loss function
l
s
and
l
c
to transfer knowledge from a network to another”).
It would have been obvious to a person of ordinary skill in the art, before the effective filing date of the invention, to have modified the federated learning method of the proposed combination to include training the model under a distance constraint condition comprising a constraint formed by a similarity as taught by He. The motivation to do so would have been to improve the models’ feature extraction capability and boost performance of the client and server models (He, Page 5, Paragraph 2, “Intuitively, the KL divergence loss attempts to bring the soft label and the ground truth close to each other. In doing so, the server model absorbs the knowledge gained from each of the edge models. Similarly, the edge models attempt to bring their predictions closer to the server model’s prediction and thereby absorb the server model knowledge to improve their feature extraction capability”; He, Page 2, Paragraph 2, Lines 8-10, “Under this reformulation, FedGKT not only boosts training CNNs at the edge but also contributes to the development of a new knowledge distillation (KD) paradigm, group knowledge transfer, to boost the performance of the server model”).
It is noted the claim recites alternative language and the proposed combination teaches at least one of the recited alternatives.
Regarding claim 17, the rejection of claim 16 is incorporated, and further, the proposed combination teaches perturbing, by the first server (Naseri, Pages 3-4, Section 2.3, Central Differential Privacy (CDP), Paragraph 2, Lines 1-5, “In this paper, we implement the CDP approach for FL dis cussed in [68] and [33], which is illustrated in Algorithm 2. The server clips the l2 norm of participants’ updates, then, it aggregates the clipped updates and adds Gaussian noise to the aggregate”), the at least one first model by using noise, to obtain at least one updated first model (Thapa, Page 8, Section 5.2, Paragraph 2, Line 6 and Equation 3, “Thirdly, calibrated noise is added to the average gradient:
g
~
k
,
t
←
1
n
k
∑
i
g
-
k
,
t
x
i
+
N
0
,
σ
2
C
'
2
I
”); and
training, by the first server (Naseri, Pages 3-4, Section 2.3, Central Differential Privacy (CDP), Paragraph 2, Lines 1-5, “In this paper, we implement the CDP approach for FL dis cussed in [68] and [33], which is illustrated in Algorithm 2. The server clips the l2 norm of participants’ updates, then, it aggregates the clipped updates and adds Gaussian noise to the aggregate”), the at least one updated first model under the constraints of the first sparsity constraint condition and the first distance constraint condition, to obtain the at least one trained first model (Thapa, Page 8, Section 5.2, Paragraph 2, Lines 2-5, and Equations 1 and 2, “Now, firstly, after t time, the client k, receives the gradients
d
A
k
,
t
from the server, and with this, it calculates client-side gradients
∇
l
k
(
W
k
,
i
,
t
C
)
for each of its local sample
x
i
, and
g
k
,
t
x
i
←
∇
l
k
W
k
,
i
,
t
C
. Secondly, the
l
2
-norm of each gradient is clipped according to the following equation:
g
-
k
,
t
(
x
i
)
←
g
k
,
t
x
i
/
max
1
,
g
k
,
t
x
i
2
C
'
”; He, Page 5, Lines 3-5 and Equation 6, “The server CNN absorbs the knowledge from many edges…To be more specific, in Eq. (2) and (5), we design
l
s
… as follows.
l
s
=
l
C
E
+
∑
k
=
1
K
l
K
D
z
s
,
z
c
k
=
l
C
E
+
∑
k
=
1
K
D
K
L
(
p
k
|
|
p
s
)
”; He, Page 5, Lines 6-7, “
D
K
L
is the Kullback Leibler (KL) Divergence function that serves as a term in the loss function
l
s
and
l
c
to transfer knowledge from a network to another”).
Regarding claim 18, the rejection of claim 16 is incorporated, and further, the proposed combination teaches wherein the first distance constraint condition further comprises a constraint formed by a similarity between the second model and a third model, and the third model is a model from an upstream device of the first server (Thapa, Page 6, Lines 3-4, “Afterward, the server updates its model by a weighted average of the gradients that it computes during the back-propagation on each client's smashed data”; see also Thapa, Page 5, Figure 2, Client 1 is considered to be “upstream” while client k is considered to be “downstream”; He, Page 5, Lines 3-5 and Equation 6, “The server CNN absorbs the knowledge from many edges…To be more specific, in Eq. (2) and (5), we design
l
s
… as follows.
l
s
=
l
C
E
+
∑
k
=
1
K
l
K
D
z
s
,
z
c
k
=
l
C
E
+
∑
k
=
1
K
D
K
L
(
p
k
|
|
p
s
)
”; He, Page 5, Lines 6-7, “
D
K
L
is the Kullback Leibler (KL) Divergence function that serves as a term in the loss function
l
s
and
l
c
to transfer knowledge from a network to another”; The “server CNN absorbs the knowledge from many edges”, thus, in combination with Thapa, the server CNN would absorb knowledge from upstream and downstream devices).
It is noted the claim recites alternative language and the proposed combination teaches at least one of the recited alternatives.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
V. Turina, Z. Zhang, F. Esposito and I. Matta, "Federated or Split? A Performance and Privacy Analysis of Hybrid Split and Federated Learning Architectures," 2021 IEEE 14th International Conference on Cloud Computing (CLOUD), Chicago, IL, USA, 2021, pp. 250-260, doi: 10.1109/CLOUD53861.2021.00038 teaches federated split learning where each pair of client and edge server exchanges the intermediate data of the neural network and the gradients, then at the end of an epoch a parameter server averages the edge servers’ weights.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to MOLLY CLARKE SIPPEL whose telephone number is (571)272-3270. The examiner can normally be reached Monday - Friday, 7:30 a.m. - 4:30 p.m. ET..
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Kakali Chaki can be reached at (571)272-3719. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from 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. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/M.C.S./ Examiner, Art Unit 2122
/KAKALI CHAKI/ Supervisory Patent Examiner, Art Unit 2122
Prosecution Insights