Pandorex
Patents

Samsung filing proposes an AI filter for suspicious storage commands

Published Pandorex Redaktion·2 min read
—

Ransomware can leave traces in the sequence of storage reads and writes. In application US20260350167A1, published on November 5, Samsung describes a filter that would look for suspicious patterns in that sequence. The model's output would help determine which commands are executed.

How the filter assesses individual commands

The model receives a sequence of storage commands. Features may include the time, block address, number of affected blocks and operation type. One variant records whether a write follows a read of the same block. Positional information preserves the order, and the transformer weighs relationships between commands. The same write could therefore be assessed differently depending on preceding activity. Claim 1 connects the assessment of each command to its execution. Claims 6 and 9 add a possible alarm and sequence analysis using self-attention, respectively.

What the storage layer can see

The filter looks for access patterns; the block features described do not require examining file contents. File paths and process identifiers are optional additional features when that information is available. The description gives an NVMe SSD as one example. It also allows a server implementation. Whether detection is isolated from the operating system and can still intervene on a compromised machine depends on the implementation.

Pandorex View: Blocking a write in time could prevent further data from being overwritten. In paragraph [0127], Samsung reports qualitatively that experiments outperformed conventional systems. It gives no comparative figures for detection rates, false alarms or added latency. Legitimate write-intensive tasks would be an important practical test. The document establishes neither reliable protection in everyday use nor a product on the market.

Document status, checked November 7, 2026: Samsung Electronics Co., Ltd. filed US application 19/196,188 on May 1, 2025. The document identifies no earlier priority. The November 5, 2026 A1 publication establishes that the application was disclosed. We could not independently confirm any subsequent prosecution status.

Sources and references

Sources used for the facts and context in this article.

  1. USPTO, US20260350167A1: Transformer-Based Ransomware Detection for Memory Commands; Deckblatt, Absätze [0024]–[0027], [0036], [0048]–[0057], [0114]–[0118] und [0127], Ansprüche 1–9 und 16–18image-ppubs.uspto.gov

How Pandorex researches and corrects articles

Comments

Sign in to write a comment.

Swipe up
Next Article

TSMC filing proposes a glass buffer to reduce chip package warpage

Patents