Preliminary Research Findings on the Impact of Random Projection on Biometric Recognition Performance

As part of the PatentReview project, we do not limit our work to analysing patent documents. We also conduct our own technical research to better understand the practical characteristics of the solutions described in the patents under review.

As a first step, we used a large publicly available face recognition dataset containing images of more than 8,000 individuals. On average, more than 300 photographs were available for each person, enabling us to perform millions of biometric comparisons during our experiments.

During preprocessing, every facial image was normalised and converted into an embedding vector using a state-of-the-art deep neural network. These vectors provide a mathematical representation of the most distinctive facial characteristics and form the basis of modern face recognition systems.

The embedding vectors were then projected into a lower-dimensional space using random Gaussian projection. This transformation alone resulted in a measurable reduction in recognition performance. Under the configurations we examined, the system achieved a True Acceptance Rate (TAR) of approximately 70–79% at a fixed False Acceptance Rate (FAR). Recognition performance deteriorated further as the output dimensionality was reduced.

We subsequently performed a second experiment that more closely reflects the approach described in the patents under review. In this simulation, a new randomly generated Gaussian projection was used for every individual authentication transaction. The comparison metric and decision threshold determined during the first experiment remained unchanged, meaning that the projection strategy was the only variable.

According to our preliminary results, authentication performance deteriorated significantly under these conditions, to the extent that, for the tested configuration, the practical usability of the system became highly questionable.

It is important to emphasise that these findings are preliminary research results. A single simulation is not sufficient to draw general conclusions regarding either the technical merits or the practical applicability of a patented invention.

Our next objective is to repeat these experiments using multiple publicly available datasets, different neural network architectures, alternative projection methods, and a broader range of parameter settings. We also intend to seek independent verification of our findings through collaboration with external researchers and domain experts.

Should these additional studies produce comparable results, they may provide important evidence when assessing the technical effect achieved by the solution described in the patents and, consequently, its relevance to the legal assessment of patentability.

Why did we use face recognition?

Our experiments were conducted using a face recognition dataset because the claims of the patents under review are not limited to a single biometric modality. Rather, the claimed scope generally refers to the use of biometric data, which may encompass a variety of biometric characteristics, including facial images, fingerprints, iris patterns, or other biometric identifiers, depending on the interpretation of the claims.

For this reason, we believe it is technically appropriate to evaluate the proposed approach using a biometric modality for which large, publicly available scientific datasets exist. Face recognition is particularly well suited for this purpose because it enables reproducible experiments and objective comparison of results across different implementations.

Naturally, this does not imply that results obtained using face recognition automatically apply to every biometric technology. PatentReview does not seek to draw universal conclusions from a single experimental study. Rather, our objective is to investigate how the technical principles described in the patents may affect the performance of different biometric systems.

In later phases of the project, we intend to extend our research to additional biometric modalities.