Deepfake Forensics Workflow for Video Analysis

Reading time: 8 min

This guide takes you through a forensic video analysis workflow covering visual artifacts, metadata, AI-based detection, geometric consistency and spectral analysis. It helps you cross-check findings and reach conclusions that are explainable, reproducible and forensically defensible.

Determining whether a video is authentic or manipulated is one of the most pressing challenges in modern forensic media analysis. Public discussion often oversimplifies this issue to a binary question: “Is this a deepfake?”. In reality, true forensic analysis requires a much deeper investigation.

Which anomalies should we look for, and where do we actually begin when a video’s integrity is called into question?

In this post, we outline a scientific workflow for forensically sound video analysis. This methodology is designed not just to detect manipulation, but to generate evidence through an approach that is explainable, reproducible, and legally defensible.

Prerequisite: This guide is written for forensic practitioners who are already familiar with the fundamentals of digital imaging and core forensic methodologies. If you need a refresher, read our overview of the fundamental principles and challenges of working with image and video evidence.

Key Takeaways

• No single test can tell the whole video story.
• Video analysis provides temporal clues that are not present in a single image, including flickering, warping and frame-to-frame inconsistencies.
• Metadata can reveal useful clues, but missing or unusual fields must always be interpreted in context.
• AI-based detectors can help locate suspicious content, but their results should guide and support the investigation rather than replace it.
• Traditional forensic methods remain essential because many manipulations can be performed without generative AI.
• The strongest conclusions come from independent findings that support one another across visual, structural, geometric and spectral analysis.

How to Analyze a Video for Signs of Deepfake Manipulation

The First Step: Visual Analysis

The most intuitive place to start is visual analysis. Generative AI models are notorious for introducing physical anomalies, particularly when rendering complex details like human hands, facial features, text, and fine patterns.

In a static image, an artifact is either present or absent. In video, however, anomalies often materialize only briefly, as frame-by-frame inconsistencies, often appearing as a sudden flickering effect. In the example below, look closely at how the subject’s thumbnail warps awkwardly for just a single frame.

We can also expose these anomalies by analyzing the video’s spatial continuity. For instance, in this next example, notice how the “STOP” sign suddenly shifts its appearance and orientation between two consecutive frames.

While visual inspection is a vital starting point, it is rarely straightforward in practice. In real-world forensic cases, source videos are often heavily compressed. Distinguishing between a synthetic “AI-generation artifact” and a standard, expected “compression artifact” can be incredibly difficult, even for experienced examiners. Moreover, as generative models continue to rapidly improve, these visual cues will only become scarcer and harder to spot.

Examine the Video Metadata and File Structure

The video stream itself is only one piece of the puzzle. It resides within a digital container file packed with specific encoding properties and metadata. Naturally, the next step is to scrutinize the file structure and metadata for any hidden AI footprints.

Even when a video’s visual content appears flawless, the underlying metadata can tell a different story. However, interpreting this data requires caution, as certain anomalies can easily mislead an examiner. Take, for example, the metadata extracted from the “broken smartphone” video below:

Media Create Date: 0000:00:00 00:00:00
Media Modify Date: 0000:00:00 00:00:00

While these fields should ideally contain valid timestamps, seeing null or “zeroed-out” values is not automatically proof of tampering. Metadata population varies wildly across different device brands, operating systems, and camera models.

Fortunately, some generative tools actively leave clear indicators in the file structure, if you know where to look. In the very same “broken smartphone” file, for instance, we can locate an explicit AIGC (Artificial Intelligence Generated Content) tag embedded in the metadata.

Note that while finding an AIGC tag is highly significant, it rarely explains the complete history of the media on its own. The tag itself is compatible with several distinct scenarios: the video could be entirely AI-generated, partially manipulated in a specific region, or simply processed or exported using an AI-assisted editing tool.

Use AI-Based Detection to Identify Suspicious Frames

Another critical piece of the puzzle is using AI to detect AI. Just like we do in our image-based workflows, we can leverage deep learning models to determine if a video’s pixel statistics are consistent with synthetic generation.

By applying state-of-the-art image deepfake detection algorithms frame-by-frame across a video, we can isolate exactly which parts of a timeline are suspicious. In the example below, we ran a questioned video through the deepfake detection filters in Amped Authenticate’s Video Mode. The analysis revealed a highly consistent compatibility score (averaging over 80%) with known AI generation models across the flagged frames.

While this automated screening is incredibly powerful, we must remember that deep learning tools are not a silver bullet for forensic investigations. On their own, “black box” neural network outputs struggle to meet strict legal and forensic standards. Rather than serving as final proof, AI-based detection is best used as a triage tool to flag anomalies and guide deeper into more explainable analyses.

Assess Frames’ Geometric-based Consistency

This is where checking a video’s physical properties becomes invaluable. Over the last few years, forensic research has shown that AI-generation models may struggle with physical consistency by introducing geometric discrepancies in shadows, perspective lines, and reflections¹.

While these subtle physical inconsistencies can easily trick the human eye, they can be mathematically analyzed and verified using geometric forensic tools (learn more in our blog post on shadow and reflection analysis). Furthermore, because we are dealing with video, maintaining this physical consistency over time across multiple frames is an incredibly complex task for generative algorithms.

A classic, pre-AI example perfectly illustrates this vulnerability. Years ago, a viral video showing a golden eagle snatching a child in a park fooled millions of viewers and was even featured on major networks like ABC News. While the visual quality easily deceived the public, a geometric analysis of the shadows cast by the eagle’s wings quickly exposed the hoax. The angle of the light source required to produce the bird’s shadow simply did not match the lighting of the rest of the scene, mathematically proving the video was not authentic. If you are curious about how it works, you can find a more technical description of shadow analysis in this blog post on the subject.

It is worth noting that this viral eagle video could never have been caught by a “deepfake detector” because it was created using older, manual CGI compositing techniques rather than modern generative AI. This distinction is crucial since, when analyzing digital media, we should avoid the dangerous logical fallacy: “If a deepfake detector doesn’t flag it, the video is  authentic.”

A video can be manipulated in dozens of ways unrelated to artificial intelligence.

This is why traditional forensic tools, such as geometric and physical analysis, offer a much broader range of applications than deep learning classifiers. The trade-off is that these tools cannot tell you how an inconsistency was made. They cannot distinguish between an image created through highly skilled manual editing and one generated entirely by AI.

Fortunately, in the vast majority of forensic cases, the primary objective is to evaluate the video’s trustworthiness. Pinpointing the exact underlying technology used to alter the scene is often secondary to simply proving that an alteration took place.

Look for Hidden Clues Through Spectral Analysis

As a final step in a comprehensive multi-domain video analysis, we should not forget to look beyond what is visible to the naked eye. Academic research has repeatedly highlighted that even visually flawless AI-generated and manipulated content often exposes distinct anomalies in its spectral representation².

A perfect real-world example of this occurred during an episode of the Amped Podcast, where the hosts featured a deepfake voice clone of Arnold Schwarzenegger speaking with Amped Software’s  Forensic Director.

If you analyze the audio file using a Fourier transform and look at the spectrogram amplitudes, the deception becomes visible as an unexpected frequency peak spike during “Arnold’s” synthesized speech segments.

The voice clone was, of course, created as a joke. However, it also provided a practical demonstration of how spectral analysis can reveal signs of synthetic manipulation. Watch the full Amped Podcast episode to see the example in context.

Conclusions

The described workflow is built on a fundamental principle of forensic analysis: while it is relatively easy for a manipulator to conceal a single trace of alteration, producing a fake video that remains perfectly consistent across every aspect is far more difficult.

This is why the proposed methodology focuses on gathering and cross-referencing clues across multiple domains, including visual artifacts, container structure, metadata, AI-based detectors, geometric consistency, and spectral analysis. Combining evidence from these different domains supports a more solid and forensically defensible conclusion.

Ultimately, this workflow highlights that answering the seemingly simple question, “Is this video a deepfake?” requires far more than a single click on a deepfake detector. It demands rigorous investigative skills, an objective scientific methodology, and a deep, fundamental understanding of image encoding, compression, and physics.

---

1. Bianchi, Stefano, et al. “Assessing shadows and reflections consistency in AI-generated images“. 2024 IEEE International Workshop on Information Forensics and Security (WIFS). IEEE, 2024. ↩︎
2. Corvi, Riccardo, et al. “On the detection of synthetic images generated by diffusion models“. ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023. ↩︎

---

FAQ – Deepfake Forensics Workflow for Video Analysis