A Hacker News post reveals that Samsung devices running Tizen are executing a TensorFlow Lite model on 960×540 images with INT8 quantization every few seconds. The model, named "alex.tflite," appears to be a variant of AlexNet based on torchvision.models.alexnet. It performs nearest neighbor feature comparisons against a database, though it reportedly fails to function correctly 99% of the time.
The findings suggest the on-device inference wastes CPU resources without delivering meaningful results. The consent gate appears to be applied after processing, meaning data may be processed before user consent is obtained. Researchers have shared the TensorFlow model graph and captured network traffic, providing a detailed breakdown of Tizen-related privacy and performance concerns.
The discovery highlights potential inefficiencies and privacy issues in Samsung's on-device machine learning implementations.
Source: Hacker News · Summarized by HeadlinesBriefing