Analysis and evaluation of autoencoder-driven dimensionality reduction for face recognition pipelines

The artificial intelligence landscape provides an ever-growing range of technologies for biometric identification, many of which use facial recognition. Usually, it conveys the usage of a database connection for storing and matching facial features, which raises privacy and security concerns. A novel Autoencoder-Driven Dimensionality Reduction ( A D D R ) architecture is proposed that enables connectionless biometric identification pipelines using a low-density QR code. It allows direct facial validation as a self-contained asset without storage or connection requirements. ADDR has been tested on top of Google’s FaceNet model embeddings and reconstructs them with a minimal loss. The accuracy in the reconstructions lets us classify the faces using the same metrics as the ones used in the original FaceNet model. At the same time, the embeddings’ compactness allows them to be stored in low-density QR codes that standard cameras can process. Several autoencoder strategies and loss functions have been designed and tested using open-source facial datasets, considering the underlying geometry of the embeddings. Also, an efficient QR encoding mechanism is defined and tested to increase the compression capabilities. The iVAA(V) (Inter-Vector Angular Alignment loss with V dimensionality) is demonstrated as the best-performing ADDR configuration. Using a dimensionality reduction of 128/16, it achieved an AUC of 0.956, which is close to FaceNet’s performance and improves the results of other dimensionality reduction techniques. • Autoencoder-Driven Dimensionality Reduction for connectionless biometrics via low-density QR. • 98% compression with 89.5% TP & 10.2% FP in facial validation tests. • Autoencoders & heuristics for high-efficiency QR-based facial encoding. • Secure authentication for tickets, transport, healthcare, and logistics. • Self-Contained, no connectivity or storage needed — instant validation & standalone.