Deepfake Detection Listed in NIST Computer Forensics Catalog
Sensity AI’s Deepfake Detection solution is now listed in the NIST Computer Forensics Tools & Techniques Catalog, maintained by the US federal government.
The listing comes as there’s no longer a question over whether law enforcement needs AI-based analytical tools to combat the threat of generative AI. In its recent AI Strategy paper, UK policing body the NPCC (National Police Chiefs’ Council) said: “AI has the potential to make UK policing far more effective at identifying, investigating, and detecting crime,” specifically identifying deepfake detection as a priority. Meanwhile, the FBI acknowledges that “AI gives the FBI new tools and capabilities”, and said it “uses AI to sift through large amounts of data to conduct video analytics”.
But as digital forensics—an essential discipline that deals with methodologies, techniques, and tools used in the identification, collection, preservation, and analysis of digital evidence—has expanded into a multifaceted field, investigators need a broader range of tools to deal with different kinds of digital crime, as well as to preserve evidence throughout the investigation lifecycle.
This begs the question:
How do you identify the best AI-powered computer forensics investigation tools?
Sensity AI Deepfake Detection is now listed on the NIST (National Institute of Standards and Technology) Computer Forensics Tools & Techniques Catalog, a database of forensic software maintained by high-level federal US authorities.
Listing a total of only 18 solutions for Image Analysis (Video & Graphics Files), the NIST Tool Catalog provides an easily searchable database of forensic tools, enabling practitioners to find software or techniques that meet their specific technical needs. It provides the ability to search by technical parameters based on specific digital forensics functions.
For law enforcement specialists and forensic investigators looking for new tools, one of the benefits of the NIST library is uniform comparisons, where the catalog uses a standard taxonomy, which lets users easily compare features between solutions.
In terms of the documented technical parameters tracked by NIST, Sensity AI Deepfake Detection is one of the few solutions listed that supports metadata analysis, illicit and altered image analysis, and video content analysis.
- Metadata analysis
- Support for EXIF metadata
- Support for IPTC metadata
- Support for XMP metadata
- Detecting illicit images
- Support for detecting illicit images
- Detecting altered images
- Support for detecting altered images
- Video content analysis
- Support for video content analysis
Using analytical AI to combat generative AI in computer forensics
For the forensic community, the initialization ‘AI’ carries some problematic baggage. Often associated with generative AI, artificial intelligence is typically used to create new content or manipulate existing content. But while the output can be convincing to the human eye, it does so in an unpredictable manner with a probabilistic output. Generative AI gives rise to the multitude of threats we see today, from deepfakes featuring synthetic identities or voices, to manipulated media like face swaps and lip syncs.
On the contrary, analytical, or discriminative AI, when leveraged by AI-powered forensic tools, can detect generative patterns in synthetic or manipulated content and is focused on providing evidence-grounded output. The aim is to open up the ‘black box’, the deep learning phenomenon in which it is not clear how an analytical model arrived at its conclusions.
The stance is that you cannot put a black box AI model in the witness stand, therefore any expert who relies on the findings of analytical AI models to influence their opinion must be able to explain and defend their conclusion to maintain accountability.
Ultimately, when a forensic instrument is AI-powered, investigators and practitioners need to apply the same standards that they do to any other forensic instrument, where analysis—delivered by AI or not—is intended to support, not replace, expert judgment, and findings must be interpreted in conjunction with contextual information.
Requirements of AI-powered image analysis tools in forensic investigations
Now that we are seeing growing acceptance of AI detection tools in forensic investigations, what are the key requirements forensic practitioners and law enforcement specialists should look for?
Integrity
Any analytical AI tool used in a forensic context should be able to demonstrate integrity. That means cryptographic verification of the submitted file, a documented chain of custody, an audit trail, and an explicit declaration that no modifications were made to the original media during analysis.
Transparency
Detection models rely on algorithms that require training. So, the training data should be documented, alongside evaluation criteria for accuracy claims, any known biases due to limitation in training data, underlying model architecture, and algorithmic methodology.
While it’s true that all of this information is usually proprietary and is unlikely to be disclosed in public, it needs to be available for disclosure in court should the need arise.
Explainability
This refers to opening the black box. The methodology underpinning the analytical outputs should be backed by peer-reviewed literature, and decisions should be explained where possible.
Accountability
It’s tempting to envision a one-click-decision-tool that tells you if an image is fake or not. But machines cannot take accountability and therefore it always comes down to the human expert who takes responsibility for the interpretation and presentation of any results. The tool produces the analysis, but the practitioner draws the conclusion.
That is the correct forensic relationship between instrument and examiner – no different from any other forensic discipline.
