Samsung running a TensorFlow lite model on 960x540 image to track you

What happened

A developer named mike64 looked at a Samsung device and found a file called alex.tflite. It is a small AI model built on AlexNet, an older image-recognition design. The device appears to take screenshots of the screen, about 960 by 540 pixels, and run the model on them about every two seconds. The goal seems to be figuring out which game is being played. The developer says the results go nowhere useful yet, because the consent step comes later in the process.

Key facts

• Image size checked: 960 x 540 pixels
• Run rate: about every couple of seconds
• Model file: alex.tflite, based on AlexNet
• Number format: INT8, a compressed form that uses 8-bit whole numbers
• Developer's view: the matching system often fails, about 99% of the time

Some plain-language terms

TensorFlow Lite is a small version of Google's AI tool that runs on phones and devices. AlexNet is an older model that learns to recognize pictures. INT8 quantization makes a model smaller and faster by using simpler numbers. The developer's findings come from reading the device's files and watching what it does, not from an official Samsung statement.

Why it matters

Many people do not know their TV may be watching what is on screen. Running AI on screen images all the time uses power and processing. It also raises privacy questions, especially if this happens before a user says yes. The developer calls this a strong example of problems in Tizen, Samsung's TV software, seen on a real device.

What to watch

• Whether Samsung confirms or explains the model
• Whether the consent prompt is moved earlier or the feature is turned off
• Further checks from the developer and other researchers
• Any change to how the TV handles screen data