Samsung Patents a Way to Shield On-Device AI Data During Loading
Every time a phone runs an AI task, it has to move a huge pile of data from storage to working memory. Samsung has patented a method that keeps that data scrambled during the trip, so it's never exposed in a readable form in between.
How Samsung protects AI data moving between memory chips
Every time your phone runs an AI feature, say, rewriting a text message or enhancing a photo, it pulls a bundle of learned instructions called weights out of long-term storage and into faster, temporary memory so the processor can use them. That transfer is a moment of vulnerability: anything sitting in intermediate memory could, in theory, be read by malicious software.
Samsung's patent describes a two-stop relay system. The weights get compressed (scrambled into a compact form) before they leave permanent storage, travel to a first temporary memory chip in that protected state, and only get decompressed into their usable form once they reach a second, more secure temporary memory chip right before the processor needs them.
The result is that the weights are never sitting in plain, readable form at an exposed waypoint. If something intercepts the data mid-journey, it gets gibberish. It's a straightforward idea borrowed from how encrypted file transfers work on the internet, applied to the internal highway inside your phone.
… compress, through the compressor, partial weights related to the specified operation, wherein the partial weights correspond to some of the weights among the weights stored in the non-volatile memory, and load the compressed partial weights into the first volatile memory …
Translation: It shrinks only the specific pieces of the AI model needed for a task before moving them.
How the two-stage compress-then-decompress transfer works
The patent describes a pipeline with three memory tiers and two processing steps.
Non-volatile memory (think: the phone's permanent storage, like flash memory) holds the full set of AI model weights. When the device needs to run a specific AI operation, it doesn't load all the weights at once. It loads only the partial weights relevant to that task, a design that saves memory bandwidth and limits exposure.
Before those partial weights leave non-volatile storage, a dedicated hardware block called the compressor encodes them using a compression algorithm. The compressed weights land in the first volatile memory (fast, temporary RAM), where they remain in their compact, opaque form. A second hardware block, the decompressor, then decodes them and loads the readable weights into a second volatile memory, which sits closer to the processor and is presumably more tightly access-controlled.
The processor performs the AI operation using the decompressed weights from that second memory only. At no point are the weights in a human-readable state at an intermediate, less-protected stop. The claim also specifies that only partial weights tied to the requested operation are moved, so the full model is never fully exposed at once:
- Weights stored compressed in non-volatile memory
- Compressed partial weights loaded to first volatile memory
- Decompression occurs, results go to second volatile memory
- Processor reads only from the second, decompressed store
The electronic device may decompress the some weights compressed and loaded to the first volatile memory by means of a decompressor and load the decompressed weights to a second volatile memory.
Translation: The device expands those compressed pieces in a secure workspace right before using them.
What this means for AI privacy on your Samsung device
On-device AI is only useful if users can trust that the AI model itself, which a company may have spent millions training, isn't trivially copyable or readable by a malicious app. This patent addresses model theft and snooping at the hardware level, not just through software locks. That matters more as phones take on more sensitive AI workloads like health data analysis or private document summarization.
From a practical standpoint, this architecture could appear in any Samsung device running local AI. Samsung keeps filing on on-device AI security The two-memory-tier design also has a secondary benefit: loading only the weights needed for a given operation at a time is more memory-efficient, which matters on a phone where RAM is always tight.
Samsung's 23rd filing we've tracked since June adds to a pattern that includes letting small models teach large ones and compressing models without losing accuracy, all part of our on-device AI privacy watch.
Claim 1 is written broadly. It covers any electronic device with a compressor, decompressor, non-volatile memory, and two volatile memories arranged in this compress-transfer-decompress sequence, without specifying the compression algorithm, the memory technology, or the AI model type. That breadth is a double-edged situation for Samsung.
On one hand, a wide claim could block competitors from implementing any similar two-stage memory protection scheme for AI weights on a device, regardless of how they implement the compression. On the other hand, wide claims are exactly what patent examiners scrutinize hardest. The basic idea of compressing data before moving it between memory tiers is well-established in computing, and Samsung will need to argue convincingly that applying it specifically to AI model weights in this exact architecture is non-obvious.
The filing is practical rather than spectacular. It addresses a real problem (protecting model weights during in-device transfer) with a logical solution. Whether it survives examination in its current broad form is an open question, but the underlying engineering approach is sound and the security motivation is clear.
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
15 drawing sheets from US 2026/0291516 A1 · click any drawing to enlarge
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →
Be the first to weigh in