# Integrating 1D-CNN and Bi-GRU for ENF-Based Video Tampering Detection

## TL;DR

This paper introduces a new method using neural networks to detect video tampering by analyzing power grid frequency signals in videos, without needing a reference database.

## Contribution

The first data-driven approach for ENF-based video forensics without relying on ground-truth ENF databases.

## Key findings

- A preprocessing stage was proposed to reduce the impact of moving objects on luminance signals.
- An anomaly detection model combining 1D-CNN and Bi-GRU achieved effective inter-frame tampering detection.
- The method was tested on both static and dynamic video datasets, showing promising results.

## Abstract

Electric network frequency (ENF) refers to the transmission frequency of a power grid, which fluctuates around 50 Hz or 60 Hz. Videos captured in a power grid environment may exhibit flickering artifact caused by the intensity variation in the light source, thus exhibiting the flickering pattern according to the ENF fluctuation. This flicker, notable for its temporal dynamics and quasi-periodic property, acts as an effective means for video tampering forensics. However, ground-truth ENF databases are often unavailable in a real-world authentication setting, thus posing challenges in conducting ENF examination in video forensics. In addition, dynamic scenes in videos also increase the difficulty of anomaly detection in ENF signals. To address these challenges, we proposed an approach based on neural networks to detect inter-frame tampering in CMOS videos that incorporate ENF signals. To the best of our knowledge, this is the first work that deploys data-driven approach for ENF-based video forensics. Without the aid of the reference ENF dataset, we exploited the implicit ENF variation in luminance signals and transformed the video signal into a one-dimensional time series utilizing ENF priors. In addition, to alleviate the impact of moving objects that also cause the variation in luminance signal, a preprocessing stage is proposed. On this basis, we designed an anomaly detection model combining 1D-CNN and Bi-GRU to conduct experiments on static and dynamic video datasets. The experimental results demonstrate the effectiveness of our proposed method in inter-frame video tampering detection, implying its potential as a forensic tool for ENF-based video analysis.