Samsung · AI / ML · Filed May 27, 2026 · USPTO published Nov 12, 2026 · This week's release

A New Samsung Patent Tracks Which Phone Sent a Bad AI Update

Phones that tweak a shared AI to fit you can also spoil it. Samsung's patent application gives every tuned model a signed tag tied to the chip that made it, so the server can spot and shut out repeat offenders.

By Patentlyze Team · Updated Nov 12, 2026 · 3 min read

A smartphone improves a shared AI using personal data and verifies its identity when sending improvements back to a server. Drawing from patent filing US 2026/0359036 A1. — A smartphone improves a shared AI using personal data and verifies its identity when sending improvements back to a server.All 14 drawings ↓

Publication number: US 2026/0359036 A1(US20260359036A1)
Applicant: SAMSUNG ELECTRONICS CO., LTD.
Filed · Published: May 27, 2026 · Nov 12, 2026
Inventors: Kyungsun CHO, Jaemin RYU
Status when we published: Waiting for an examiner
Document: 20 claims
US classification: 717/168
Parent application: Continuation of PCT/KR2026/096855 (filed May 12, 2026)(PCTKR2026096855)

What Samsung's signed personal AI models do

Your phone's AI can learn from how you use it, and a company can fold those lessons back into the shared AI everyone downloads. The catch is that some of what phones learn is biased, distorted, or planted on purpose, and a shared AI that absorbs it gets worse for everybody.

Samsung's patent application describes a paper trail for that process. After your phone tunes its own copy of the AI, it stamps the result with an ID built from tiny physical quirks of its processor chip, plus details like the region and time. It signs the model with that ID and sends both to Samsung's server.

If the server finds a model that has drifted too far from the original, it can remember where that model came from. Later models from the same phone, from other phones owned by the same person, or from the same group of devices can then be turned away.

From the filing · CLAIM 1

… obtaining, by the user terminal, a personalized AI model by performing fine-tuning on the base AI model based on user data, wherein the user data is collected in a process of executing the base AI model on the user terminal; …

Translation: The phone customizes its AI software by learning from how you use it daily.

How the server spots and blocks bad AI models

The phone first downloads a base AI model (the general-purpose version) from the server. It then fine-tunes it (retrains it slightly) on what you type, say or photograph while using it. The filing says this happens in a trusted execution environment (a walled-off part of the chip that ordinary apps cannot touch), and the input is anonymized or reworded first, for example by swapping synonyms, to protect privacy.

Next the phone builds model-specific information. The core is a key from a physical unclonable function (a fingerprint made from microscopic manufacturing differences in the chip). Around it go:

The region and locale of the phone
A group the device belongs to
The device model, plus the date and time
IDs for the secure environment that ran the tuning

Private fields are scrambled with a hash (a one-way scramble) and a random number. The phone then signs the model's structure and values with the key, and sends everything to the server.

On the server, the check is numeric. It compares each setting in the personalized model with the base model, then looks at the average gap, the spread, or how many settings moved past a limit. In the filing's example, two settings shifting by more than 0.4 get a model flagged. Once a model is flagged, later models from the same chip, the same owner, or the same group can be rejected or re-checked.

Why Samsung wants to trace bad AI updates to a phone

If your phone ever lets an AI learn your habits and then shares what it learned, someone has to decide whose lessons count. This filing gives the server a way to say no, and to remember who it said no to. A bad batch from one device could be kept out of the update that reaches everyone else.

There is a trade-off. Tying every submission to a specific chip means the server can recognize a repeat sender, though the filing scrambles the private chip ID fields. And excluding by owner or group means one flagged device could cut off others. How fair that turns out depends on how a company applies it.

Samsung's 15th filing we've tracked since June in our AI guardrails race watch builds on earlier applications like one blocking data misuse and one curbing photo hallucinations.

Editorial take

The problem is real. An AI that updates itself from millions of users' habits is a tempting target, and one corrupted batch can leave the shared version a little worse for every person who downloads it. Fixing a spoiled model after the fact costs far more than screening it on the way in.

Samsung's answer fits that size of problem reasonably well. Measuring how far a returning model drifted is a simple, cheap test, and tying each model to a specific chip makes it hard for a troublemaker to keep resubmitting. The filing does not say how the cutoffs get chosen, and that decides whether the screen catches real damage or just ordinary personal quirks.

The group-based blocking is the shakiest part. Shutting out everyone in a group because one device misbehaved could punish innocent people, and the filing gives little sense of how that risk is weighed.

Source: US 2026/0359036 A1 at the USPTO (PDF). Full patent text and figures are in the official publication.

Get our take in your Top Stories

Liked this breakdown? Add Patentlyze as a preferred source on Google, and our plain-English take shows up more often in your Top Stories the next time Samsung patent news breaks.

Add Patentlyze on Google One click.

There are more where this came from

We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.

The drawings · 14 sheets

14 drawing sheets from US 2026/0359036 A1 · click any drawing to enlarge

How phones train AI locally and update the cloud, drawing sheet 1 from US 2026/0359036 A1 — Sheet 1 How phones train AI locally and update the cloud

Isolating user training data inside a secure phone chip, drawing sheet 2 from US 2026/0359036 A1 — Sheet 2 Isolating user training data inside a secure phone chip

Steps for swapping AI training data with the server, drawing sheet 3 from US 2026/0359036 A1 — Sheet 3 Steps for swapping AI training data with the server

Signing phone AI updates with a unique chip ID, drawing sheet 4 from US 2026/0359036 A1 — Sheet 4 Signing phone AI updates with a unique chip ID

Scrambling private phone identifiers while keeping region data intact, drawing sheet 5 from US 2026/0359036 A1 — Sheet 5 Scrambling private phone identifiers while keeping region data intact

Checking if an incoming phone update looks abnormal, drawing sheet 6 from US 2026/0359036 A1 — Sheet 6 Checking if an incoming phone update looks abnormal

Testing a test model update to spot harmful changes, drawing sheet 7 from US 2026/0359036 A1 — Sheet 7 Testing a test model update to spot harmful changes

Blocking future updates from a compromised phone, drawing sheet 8 from US 2026/0359036 A1 — Sheet 8 Blocking future updates from a compromised phone

Show the other 6 sheets

Blocking updates from phones tied to a bad source, drawing sheet 9 from US 2026/0359036 A1 — Sheet 9 Blocking updates from phones tied to a bad source

Main internal hardware parts running inside the user phone, drawing sheet 10 from US 2026/0359036 A1 — Sheet 10 Main internal hardware parts running inside the user phone

How a phone tunes and signs an AI model, drawing sheet 11 from US 2026/0359036 A1 — Sheet 11 How a phone tunes and signs an AI model

How the server screens incoming updates against past bad sources, drawing sheet 12 from US 2026/0359036 A1 — Sheet 12 How the server screens incoming updates against past bad sources

Creating device identifiers from a chip's unique physical fingerprint, drawing sheet 13 from US 2026/0359036 A1 — Sheet 13 Creating device identifiers from a chip's unique physical fingerprint

Signing trained AI files using the hardware-derived security key, drawing sheet 14 from US 2026/0359036 A1 — Sheet 14 Signing trained AI files using the hardware-derived security key

Watchlists this filing belongs to

All watchlists →

The AI guardrails raceCross-company · 228 filings

Reader comments

The drawings · 14 sheets

More from Samsung

Samsung Files Patent for One AI Model That Handles Many Editing Jobs

Samsung Files Patent to Turn Your Photo Into Custom App Icons

New Patent Reads Phone Logs to Infer User Interests

Samsung Files Patent to Help AI Learn New Things Without Forgetting Old Ones

More in AI / ML

Google Patents Vibrations That Tell You How Sure Its AI Is

IBM Files Patent for Showing Which Words or Image Areas Swayed an AI

Microsoft Files Patent for AI Coding Help That Checks Its Own Work First

IBM Files Patent to Stop AI From Being Fooled by the Names in Your Questions

Watchlists this filing belongs to

No comments yet

Start the discussion Cancel reply