Granted Patent B1

US 12,463,947 · App. 18/333,482 · Granted Nov 4, 2025

Inventors: Eric Crockett (Redmond, WA); Gang Wang (Frederick, MD); Joan Feigenbaum (New York, NY)

Assignee: Amazon Technologies, Inc.

H04L63/0428G06F21/6254G06F21/6263

Abstract

Claims (68)

1 . A system, comprising:

one or more processors with associated memory that implement a privacy preserving supplemental content server (PPSCS), configured to:

serve supplemental content to a plurality of users when the users consume content published by a content server, wherein the supplemental content are selected based at least in part on user-private information about the users;

store a first secret share of a user-to-segment map (USM), wherein the USM contains the user-private information about the users, wherein the first secret share does not reveal the user-private information, and wherein a second secret share of the USM is stored at the content server;

receive a request from the content server to serve supplemental content to a user, wherein the request indicates a request context determined by the content server;

execute one side of a privacy preserving protocol with the content server to perform a lookup of user segment data of the user in the USM based at least in part on the request context, wherein the privacy preserving protocol guarantees that:

(a) the content server does not learn any user-private information about the user in the USM,

(b) the PPSCS learns at most one user segment of the user, and

(c) the PPSCS cannot track the user over multiple requests from the content server using any user-private information about the user in the USM learned during the execution;

select the supplemental content for the request based at least in part on results of the lookup; and

send the supplemental content in a response according to the request, wherein the response does not reveal the supplemental content to the content server.

2 . The system of claim 1 , wherein the PPSCS is configured to:

cluster user segments of the users into multiple segment clusters;

determine cluster utility values for individual ones of the segment clusters based at least in part on the request context;

during the privacy preserving protocol, perform part of a two-party computation (2PC) with the content server to (a) select a key segment of the user from an optimal segment cluster determined using the cluster utility values, or (b) determine that no satisfactory key segment is found for the user in the segment clusters; and

select (a) a supplemental content specific to the user using the key segment of the user when the key segment is found, or (b) a supplemental content not specific to the user when no satisfactory key segment is found.

3 . The system of claim 2 , wherein to perform the 2PC, the PPSCS evaluates a garbled circuit generated by the content server.

4 . The system of claim 1 , wherein to execute the privacy preserving protocol, the PPSCS is configured to:

execute a part of a randomized oblivious secret share extraction (ROSSE) protocol with the content server, wherein the execution of the ROSSE protocol (a) generates fresh secret shares for one or more entries in the USM associated with the user in the USM, and (b) causes the content server and the PPSCS to learn the fresh secret shares but nothing else in the USM.

5 . The system of claim 1 , wherein the first secret share of the USM is implemented as a Cuckoo hash table.

6 . The system of claim 1 , wherein the PPSCS is implemented using compute resources provisioned in a multitenant infrastructure service provider network.

7 . A method, comprising:

performing, by a privacy preserving supplemental content server (PPSCS) implemented by one or more processors with associated memory:

storing a first secret share of a user-to-segment map (USM), wherein the USM contains user-private information used to select supplemental content for users consuming content at a content server, wherein the first secret share does not reveal the user-private information, and wherein a second secret share of the USM is stored at the content server;

receiving a request from the content server to serve supplemental content to a user, wherein the request indicates a request context determined by the content server;

executing one side of a privacy preserving protocol with the content server to perform a lookup of user segment data of the user in the USM based at least in part on the request context, wherein the privacy preserving protocol guarantees that:

(a) the content server does not learn any user-private information about the user in the USM,

(b) the PPSCS learns at most one user segment of the user, and

(c) the PPSCS cannot track the user over multiple requests from the content server using any user-private information about the user in the USM learned during the execution;

selecting the supplemental content for the request based at least in part on results of the lookup; and

sending the supplemental context in a response according to the request, wherein the response does not reveal the supplemental content to the content server.

8 . The method of claim 7 , further comprising the PPSCS:

clustering user segments of the users into multiple segment clusters;

determining cluster utility values for individual ones of the segment clusters based at least in part on the request context;

during the privacy preserving protocol, performing part of a two-party computation (2PC) with the content server to (a) select a key segment of the user from an optimal segment cluster determined using the cluster utility values, or (b) determine that no satisfactory key segment is found for the user in the segment clusters; and

selecting (a) a supplemental contest specific to the user using the key segment of the user when the key segment is found, or (b) a supplemental content not specific to the user when no satisfactory key segment is found.

9 . The method of claim 8 , further comprising the PPSCS:

encoding respective centroids of the segment clusters as cluster feature vectors;

encoding the request context is as a context feature vector; and

wherein a cluster utility value is determined as a dot product of the context feature vector and a cluster feature vector.

10 . The method of claim 8 , wherein performing the 2PC comprises evaluating a garbled circuit generated by the content server.

11 . The method of claim 8 , wherein executing the privacy preserving protocol comprises the PPSCS performing parts of two or more 2PCs with the content server.

12 . The method of claim 7 , wherein executing the privacy preserving protocol comprises the PPSCS:

executing a part of a randomized oblivious secret share extraction (ROSSE) protocol with the content server, wherein the execution of the ROSSE protocol (a) generates fresh secret shares for one or more entries in the USM associated with the user in the USM, and (b) causes the content server and the PPSCS to learn respective ones of the fresh secret shares but nothing else in the USM.

13 . The method of claim 12 , wherein execution of the ROSSE protocol comprises the PPSCS executing part of an oblivious transfer protocol with the content server.

14 . The method of claim 7 , wherein executing the privacy preserving protocol comprises the PPSCS executing a part of a Beaver Triples protocol with the content server to compute one or more dot products of vectors or matrices.

15 . The method of claim 7 , wherein the first secret share of the USM is implemented as a Cuckoo hash table.

16 . The method of claim 8 , wherein an entry in the first secret share of the USM corresponds to an individual user, and the entry includes a cluster membership bit vector of the individual user whose bits indicate whether the individual segment clusters contain a user segment associated with the individual user.

17 . The method of claim 16 , wherein:

the PPSCS receives a random bit vector generated by the content server, wherein the random bit vector is of a same length as the cluster membership bit vector and contains random bit values; and

the optimal segment cluster is determined based at least in part on an AND operation performed on the cluster membership bit vector and the random bit vector.

18 . The method of claim 7 , wherein:

the USM is constructed by an advertiser distinct from a publisher that operates the content server; and

the PPSCS is part of an ad delivery service operated by a service operator distinct from the advertiser and the publisher.

19 . One or more non-transitory computer-accessible storage media storing program instructions that when executed on one or more processors of a privacy preserving supplemental content server (PPSCS), cause the PPSCS to:

store a first secret share of a user-to-segment map (USM), wherein the USM contains user-private information used to select supplemental content for users consuming content at a content server, wherein the first secret share does not reveal the user-private information, and wherein a second secret share of the USM is stored at the content server;

receive a request from the content server to serve supplemental content to a user, wherein the request indicates a request context determined by the content server;

execute one side of a privacy preserving protocol with the content server to perform a lookup of user segment data of the user in the USM based at least in part on the request context, wherein the privacy preserving protocol guarantees that:

(a) the content server does not learn any user-private information about the user in the USM,

(b) the PPSCS learns at most one user segment of the user, and

(c) the PPSCS cannot track the user over multiple requests from the content server using any user-private information about the user in the USM learned during the execution;

select the supplemental content for the request based at least in part on results of the lookup; and

send the supplemental content in a response according to the request, wherein the response does not reveal the supplemental content to the content server.

20 . The non-transitory computer-accessible storage media of claim 19 , wherein the program instructions when executed on the one or more processors cause the PPSCS to:

cluster user segments of the users into multiple segment clusters;

determine cluster utility values for individual ones of the segment clusters based at least in part on the request context;

during the privacy preserving protocol, perform part of a two-party computation (2PC) with the content server to (a) select a key segment of the user from an optimal segment cluster determined using the cluster utility values, or (b) determine that no satisfactory key segment is found for the user in the segment clusters; and

select (a) a supplemental content specific to the user using the key segment of the user when the key segment is found, or (b) a supplemental content not specific to the user when no satisfactory key segment is found.

Assignments (1)

References Cited (17)