# US 12,762,276 B2 · Data depersonalization using on-device generative models > Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt - Number: US 12,762,276 B2 - Document ID: US12762276B2 - Status: Granted - Kind: B2 - Application number: 18/926,061 - Filed: 2024-10-24 - Published: 2026-04-30 - Granted: 2026-10-13 - Earliest claimed: 2024-10-24 - Assignees: Google Inc. - Applicants: Google LLC - Inventors: Milen Ferev; Luke Boyer - Web page: https://patentgrep.com/patent/US12762276B2 ## CPC - G06F 21/6245: {Protecting personal data, e.g. for financial or medical purposes} - G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N) ## Abstract Implementations relate to depersonalizing and/or anonymizing generative model inputs and/or outputs to remove sensitive and/or private information, such as PII (personally identifiable information). In various implementations, a generative model input prompt may be retrieved from local memory of the edge computing device and further processed using one or more generative models to generate generative model output. The generative model input prompt may be submitted to an edge-based anonymization process including assembling, as an anonymization prompt and processing the anonymization prompt using one or more on-device generative models of the edge computing device to generate an anonymized version of the portion of the generative model input prompt that is stripped of PII. The anonymized version may further be inspected of any remaining PII and the fully anonymized version may be surfaced without concern of leaking PII. ## Claims ### Claim 1 A method implemented using one or more processors of an edge computing device, the method comprising: retrieving, from local memory of the edge computing device, a generative model input prompt that was processed using one or more generative models to generate generative model output; submitting the generative model input prompt to an edge-based anonymization process that includes: assembling, as an anonymization prompt, at least a portion of the generative model input prompt, wherein the portion comprises one or more images; processing the anonymization prompt using one or more on-device generative models of the edge computing device to generate an anonymized version of the portion of the generative model input prompt that is stripped of at least some personally identifiable information (PII), wherein at least one of the one or more on-device generative models comprises a vision language model (VLM) or an image generation model; inspecting the anonymized version for any remaining PII; and in response to a determination, based on the inspecting, that the anonymized version is free of any remaining PII, uploading the anonymized version from the edge computing device to a remote server. ### Claim 2 (depends on 1) The method of claim 1, further comprising, in response to a determination, based on the inspecting, that the anonymized version includes at least some PII, resubmitting the anonymized version to the edge-based anonymization process. ### Claim 3 (depends on 1) The method of claim 1, wherein the portion comprises a natural language query issued by a user. ### Claim 4 (depends on 1) The method of claim 1, wherein the portion comprises metadata incorporated into the generative model input prompt in conjunction with a natural language query issued by a user. ### Claim 5 (depends on 4) The method of claim 4, wherein the metadata comprises PII of the user. ### Claim 6 (depends on 1) The method of claim 1, wherein the anonymized version is inspected using a discriminator machine learning model. ### Claim 7 (depends on 6) The method of claim 6, wherein the discriminator machine learning model comprises one of the same on-device generative models, and is finetuned for detection of PII. ### Claim 8 (depends on 6) The method of claim 6, wherein the discriminator machine learning model comprises one of the same on-device generative models augmented with one or more low-rank adaptors (LoRAs). ### Claim 9 (depends on 1) The method of claim 1, wherein at least one of the one or more on-device generative models comprises a large language model (LLM). ### Claim 10 (depends on 1) The method of claim 1, wherein the image generation model comprises a diffusion model. ### Claim 11 (depends on 1) The method of claim 1, wherein one or more of the generative models that were used to process the generative model input prompt comprise at least one of the one or more on-device generative models. ### Claim 12 (depends on 1) The method of claim 1, wherein the anonymization prompt is assembled to include a command to anonymize the portion of the generative model input prompt, wherein the command to anonymize the portion of the generative model input prompt comprises a command to replace PII in the portion of the generative model input prompt with synthetic PII. ### Claim 13 (depends on 1) The method of claim 1, wherein the anonymization prompt is assembled to include a command to anonymize the portion of the generative model input prompt, wherein the command to anonymize the portion of the generative model input prompt comprises a command to replace instances of PII in the portion of the generative model input prompt with placeholders or hashed values. ### Claim 14 (depends on 1) The method of claim 1, wherein the anonymization prompt is assembled to include a command to anonymize the portion of the generative model input prompt, wherein the command to anonymize the portion of the generative model input prompt comprises a command to generate a summary of the portion of the generative model input prompt without PII. ### Claim 15 (depends on 1) The method of claim 1, wherein the edge-based anonymization process further includes assembling, into the anonymization prompt, at least a portion of the generative model output. ### Claim 16 An edge-based system comprising one or more processors and memory storing instructions that, in response to execution by the one or more processors, cause the one or more processors to: retrieve, from local memory of an edge computing device, a generative model input prompt that was processed using one or more generative models to generate generative model output; submit the generative model input prompt to an edge-based anonymization process that includes: assemble, as an anonymization prompt, at least a portion of the generative model input prompt, wherein the portion comprises one or more images; process the anonymization prompt using one or more on-device generative models of the edge computing device to generate an anonymized version of the portion of the generative model input prompt that is stripped of at least some personally identifiable information (PII), wherein at least one of the one or more on-device generative model comprises a vision language model (VLM) or an image generation model; inspect the anonymized version for any remaining PII; and in response to a determination, based on the inspecting, that the anonymized version is free of any remaining PII, upload the anonymized version from the edge computing device to a remote server. ### Claim 17 At least one non-transitory computer-readable medium comprising instructions that, when executed by one or more processors, cause the one or more processors to: retrieve, from local memory of an edge computing device, a generative model input prompt that was processed using one or more generative models to generate generative model output; submit the generative model input prompt to an edge-based anonymization process that includes: assemble, as an anonymization prompt, at least a portion of the generative model input prompt, wherein the portion comprises one or more images; process the anonymization prompt using one or more on-device generative models of the edge computing device to generate an anonymized version of the portion of the generative model input prompt that is stripped of at least some personally identifiable information (PII), wherein at least one of the on-device generative models comprises a vision language model (VLM) or an image generation model; inspect the anonymized version for any remaining PII; and in response to a determination, based on the inspecting, that the anonymized version is free of any remaining PII, upload the anonymized version from the edge computing device to a remote server.