Reading the Silicon: A Critical Visual Analysis of Hardware Fingerprints in Low-Light RAW Video

When we capture video in extreme low light, the resulting image is often a composite: half genuine optical signal, half structural illusion injected by the camera’s own hardware. This underlying structure is known as Fixed-Pattern Noise (FPN). In this paper, we present a graphical, critical analysis of the stability and origins of these hardware fingerprints using the AIM 2025 Low-Light RAW Video dataset (∼1 lux). By extracting and directly analyzing FPN heatmaps, spatial frequency charts, and hierarchical dendrograms across 14 commercial smartphone camera modules, we draw several empirical conclusions. First, FPN acts as an indelible, highly localized topographical mask that is demonstrably scene-independent. Second, sharp harmonic spikes in row wise Power Spectral Density graphs firmly implicate periodic readout electronics as the dominant source of deterministic banding. Third, the magnitude of FPN varies by up to 20× between modules housed in the same handset. Finally, hierarchical clustering reveals that FPN structural similarity does not follow simple brand-level grouping; instead, crossbrand affinities suggest that module type and the underlying sensor die matter more than manufacturer identity. Recognizing and correcting these deterministic hardware signatures is a prerequisite to meaningful low-light image enhancement.