SECURE MULTIMODAL BIOMETRIC AUTHENTICATION USING DARTS-OPTIMIZED SWIN TRANSFORMERS AND DIFFUSION-ASSISTED CANCELABLE TEMPLATE PROTECTION

Abstract

Multimodal biometric authentication is offering an effective method to identity verification by combining matching qualities of face, iris and fingerprint modalities, in real-world scenarios, but still, reaching trustworthy authentication while protecting stored biometric templates remains a noticeable hurdle; this paper indicates a secure multimodal biometric authentication framework blending Swin Transformer-based feature extraction, Differentiable Architecture Search (DARTS), reliability-aware feature fusion, and diffusion-assisted latent representation learning, in many respects. Modality-specific Swin Transformer encoders are, as a practical matter, employed to pull discriminative features from biometric inputs; indeed, what's more, until then, the network architecture is improved, within a predefined search space by dARTS. Authentication is ran using Hamming-distance-based matching between enrolled and probe templates, and diffusion-assisted learning is worked in to support latent representation robustness, followed by a secret-dependent cancelable transformation to make binary biometric templates. The pulled representations are, broadly speaking, combined through the reliability-aware fusion to create a integrated feature representation, and the proposed framework is found using precision, recall, F-Measure, and precision and compared with BiLSTM, FOA and CNN, BiLSTM, FOA approaches. The experimental results show that the proposed framework hits 94.852% precision, 98.640% recall, 96.709% F-Measure, and 98.914% accuracy, outperforming the checked strategies across all four metrics; these findings are emphasizing the potential of combining transformer-based feature extraction, architecture optimization, and multimodal feature fusion to sharpen biometric authentication performance.