Abstract
Most artificial lights fluctuate in response to the grid's alternating current, providing the potential to estimate the Electric Network Frequency (ENF) from conventional frame-based videos. Nevertheless, the performance of Video-based ENF (V-ENF) estimation largely relies on imaging quality and suffers from interference caused by non-ideal sampling, motion, and extreme lighting. We show that the ENF can be extracted without these limitations using event cameras—neuromorphic sensors that encode light intensity variations asynchronously with extremely high temporal resolution and high dynamic range. We formulate the physical mechanism for ENF captured in events and propose a robust Event-based ENF (E-ENF) estimation method. Extensive experiments on our new Event-Video ENF Dataset (EV-ENFD) demonstrate that E-ENF extracts more accurate traces, outperforming conventional V-ENF by a large margin, especially in challenging environments.
Introduction
Every power grid possesses a signature frequency (typically 50 or 60 Hz) that fluctuates slightly over time. This Electric Network Frequency (ENF) acts as a temporal "fingerprint" for multimedia forensics. Forensic experts use ENF to verify the time and location of a recording by matching the flicker captured in the media with historical grid records. However, traditional Video-based ENF (V-ENF) is fragile, often requiring static scenes, ideal lighting, and high frame rates to avoid aliasing.
The Physical Mechanism of E-ENF
The ENF signal propagates from the power supply to the scene's illumination. As shown in Figure 1, the power supply voltage $v(t)$ drives the light source, creating a flickering intensity $\ell(t)$. This light reflects off objects and is captured by a vision sensor.
Event cameras differ from traditional sensors by performing a log-transformation of light intensity. This transformation is a key feature: it preserves the ENF signal while suppressing unwanted variations in amplitude and bias. Because event cameras operate asynchronously, they provide the high temporal resolution necessary to capture high-frequency flicker without the aliasing common in standard 30 fps video.
Methodology: From Events to Frequency
The E-ENF pipeline extracts frequency data from the asynchronous event stream through three key steps:
- Uniform-Interval Temporal Sampling: Events are "binned" into uniform 1ms time slices to create a signal suitable for standard frequency analysis.
- Majority-Voting Spatial Sampling: This step is crucial for handling motion. While moving objects create bipolar events (both positive and negative changes), global light flicker creates unipolar events across the scene. By taking a majority vote of pixel polarities in each slice, the system cancels out motion noise and isolates the ENF signal.
- Harmonic Selection: The algorithm automatically selects the optimal harmonic by calculating a "smoothness score," matching the naturally slow-drifting nature of the power grid.
Experimental Results
The researchers introduced the Event-Video ENF Dataset (EV-ENFD), featuring 51 synchronized datasets across Static, Dynamic, and Extreme Lighting scenarios.
As illustrated in Figure 2, E-ENF maintains high accuracy where traditional V-ENF fails. In dynamic scenes involving handheld movement or moving objects, E-ENF achieved a Pearson Correlation Coefficient (CC) of 0.9553, compared to just 0.7538 for high-resolution video. In extreme lighting conditions, E-ENF reached a near-perfect 0.9867 CC, leveraging the sensor's high dynamic range (HDR) to see flicker in scenes that appear as black or white blobs to traditional cameras.
Conclusion
E-ENF represents a significant leap in forensic frequency extraction. By moving to a neuromorphic modality, researchers can now extract reliable ENF signals in real-world conditions—such as moving cameras or night scenes—that were previously considered impossible for forensic analysis. While event cameras are currently specialized tools, their integration into future forensic toolkits could revolutionize how we verify digital media.
Summary updated