A new approach to biometric technology is challenging the idea that faces, voices, and other identifying traits must be stored and analyzed as ordinary digital images. Researchers are exploring what are called implicit neural representations, or INRs, which describe a biometric signal as a continuous mathematical function instead of a fixed grid of pixels.
The approach was the focus of an IEEE Biometrics Council and IAPR webinar held September 24. The presentation was led by Vishal Patel, a professor of electrical and computer engineering at Johns Hopkins University.
Traditional biometric systems usually begin with sampled data. A face is stored as pixels, a voice as a sequence of audio samples, and other signals are broken into discrete measurements. Modern systems often turn those measurements into feature embeddings, which are compact numerical descriptions used for matching.
Implicit neural representations take a different approach.
Instead of storing only the sampled values, an INR uses a neural network to represent the underlying signal as a function. The network can be asked what the signal should look like at a particular coordinate or position, even between the original sample points.
That continuous representation is already being studied for images, audio, video, 3D scenes, medical imaging, radar, and compression.
A 2026 survey paper by Dhananjaya Jayasundara and Patel described INRs as a shift from discrete sampled data toward continuous functional models. The authors noted that these systems can represent many kinds of signals with the same general idea.
For biometrics, that could open several new possibilities.
One area is deepfake detection. A deepfake face may look convincing as a normal image, but the way it was generated can leave patterns that are difficult to see by examining pixels alone. Researchers are studying whether continuous neural representations can capture subtle structure that helps distinguish a real biometric sample from an artificial one.
The IEEE presentation also focused on presentation attack detection.
A presentation attack happens when someone tries to fool a biometric system using something other than a genuine live biometric sample. That can include a photograph held in front of a camera, a replayed video, a mask, or another artificial representation of a person’s face.
The National Institute of Standards and Technology defines presentation attack detection as the automated determination of whether such an attack is taking place. NIST’s current digital identity guidance requires presentation attack detection for facial recognition in certain remote identity systems.
That makes better attack detection important as face recognition spreads into phones, banking, travel, identity verification, and security systems.
INRs could give detection systems a different kind of evidence to examine.
A conventional image is tied to its pixel grid. An INR instead learns a mathematical mapping from coordinates to signal values. The weights and internal structure of that network may contain information about texture, frequency, smoothness, and other properties of the underlying signal.
Researchers can then study the representation itself rather than looking only at the reconstructed picture.
Biometric quality assessment is another possible use.
Not every face image is equally useful for recognition. Poor lighting, blur, extreme head angles, low resolution, compression, or partial obstruction can make a biometric sample harder to match. A continuous representation may offer another way to measure whether a sample contains enough reliable information for identification.
The same approach could also be used for function-space metric learning. In simpler terms, researchers can compare the learned functions representing two biometric samples rather than comparing only two images or two conventional feature vectors.
That may eventually provide new ways to decide whether two samples came from the same person.
The concept is still emerging, however, and it should not be confused with a finished replacement for today’s biometric systems.
The September 24 IEEE event was a technical webinar, not the announcement of a newly approved commercial biometric platform. The broader INR survey by Jayasundara and Patel was posted as an arXiv preprint in April, and it described the general signal-processing framework rather than proving that INRs outperform all existing biometric methods.
Several of the biometric uses discussed by IEEE remain research directions.
One of the most interesting is morph attack analysis. A face morph combines characteristics from two or more people into a single image. If the result is accepted by a face recognition system as matching both people, it can create a serious identity security problem.
NIST has studied this problem for years and has tested software designed to detect morphed face photographs. In a 2025 report, NIST said independent benchmarks show that some morph detectors have reached a level that can provide practical value, although performance still depends on the operating conditions.
Continuous representations may give researchers another tool for finding the unusual structure created when faces are digitally blended.
Privacy is another possible advantage, but it also needs careful testing.
Biometric templates are sensitive because a person cannot simply replace a face or fingerprint the way someone can change a password. Researchers are therefore interested in representations that allow a system to compare identities while reducing the amount of directly recoverable biometric information stored in a database.
IEEE’s webinar listed privacy-preserving biometrics and continuous biometric templates among future areas where INRs could be explored.
There are also technical problems to solve.
INRs can be expensive to train, and their behavior can depend heavily on network design. Researchers still need better ways to understand the information stored in network weights, improve stability, handle very large datasets, and make systems work reliably outside laboratory conditions.
The Jayasundara and Patel survey identified weight-space interpretability and large-scale generalization as open challenges.
Those issues become especially important in biometrics because errors can affect real people. A system that incorrectly accepts an attacker creates a security risk. A system that incorrectly rejects a legitimate user can deny access to an account, device, or service.
Any new biometric method therefore needs independent testing across different people, sensors, lighting conditions, attack types, and real-world environments.
For now, implicit neural representations are best viewed as a new way of describing biometric information rather than a complete new identity system. The method changes the basic object that researchers analyze, moving from pixels and fixed samples toward learned continuous functions.
That shift could prove useful because deepfakes and biometric attacks are also becoming more sophisticated.
If future studies show that INRs can expose patterns that conventional systems miss, they could become another layer in biometric security. They may also help researchers build systems that measure quality better, detect morphs, protect stored templates, and combine several biometric signals in a single framework.
The important next step will be rigorous testing. Researchers will need to show not only that continuous representations are mathematically interesting, but that they make biometric systems more accurate, secure, and reliable under real-world conditions.
IMAGE: derivative work: MaGIc2laNTern (talk) Cool_Kids_of_Death_Off_Festival_p_146.jpg: Przykuta – Cool_Kids_of_Death_Off_Festival_p_146.jpg CC3