Face Presentation Attack Detection (fPAD) is a fundamental component of secure face recognition systems. While multimodal fPAD improves robustness by exploiting complementary cues from RGB, depth, and infrared (IR) modalities, its practical deployment remains challenging when one or more modalities are unavailable or incomplete. Existing flexible-modality fPAD methods address this issue mainly through zero-filling or modality omission, which may discard useful cross-modal information and introduce modality-dependent bias. To overcome these limitations, we propose the Flexible Modality Adaptation Network (FMA-Net), a unified framework for flexible-modality fPAD that compensates for missing modalities through instruction-guided latent-space translation. Specifically, the Face Modality Instruct Translator (FMIT) estimates target modality observations from available RGB inputs, while a context-aware category enhancement mechanism embeds learnable text prompts as modality-invariant semantic anchors to alleviate modality bias. The Facial Feature Adapter (F2A) module further enhances discriminative facial representations by integrating global facial structure and local spoof-relevant cues. In addition, a Synthetic-to-Real Face Domain Alignment (SR-FDA) module reduces the feature discrepancy between translated and real modalities, with benefits that depend on the missing-modality pattern and data domain. Extensive experiments on the WMCA and PADISI-Face datasets validate the effectiveness of the proposed framework. On WMCA under intra-domain evaluation with a 70% modality missing ratio, FMA-Net achieves a mean ACER of 3.22% across four missing-modality protocols, compared with 4.24% for the strongest baseline by the same mean criterion. When both depth and IR are entirely unavailable, it achieves an ACER of 5.23%, compared with 5.99% for the best competing result in that setting. Cross-domain experiments further reveal the limitations of synthetic-to-real alignment under dataset shifts. These results demonstrate that FMA-Net provides a practical solution for fPAD under flexible and incomplete modality scenarios.
Electrocardiogram (ECG) authentication must preserve identity discrimination while limiting the consequences of template compromise. Applying protection after representation learning may alter verification utility, and low error alone does not establish security. This study evaluates application-version key-derived projection and binary quantization, training the encoder in the protected space and matching templates by average Hamming similarity. Subject- and record-disjoint experiments cover ECG-ID, Heartprint, and the PTB diagnostic database. Two separate five-seed series examine system comparisons and authentication with a unified attack matcher. The latter yields equal error rates of 7.45%, 10.65%, and 10.09%, respectively. Independently tuned comparisons and a strictly matched PTB experiment establish no consistent accuracy advantage from protected-space training. Benchmark cross-application similarity-score areas under the receiver operating characteristic curve are 0.487–0.540, versus 1.000 for unprotected embeddings; nonzero distribution-based linkability estimates rule out a claim of complete unlinkability. Unified same-version identity recovery remains 60.20–78.28%. An embedding-injection attacker given the transform and enrollment templates succeeds on 99.38–99.49% of identities within the evaluated budget. Version renewal rejects sampled stale codes, but fresh-query false acceptance reaches 1.43–25.38% at the unchanged threshold. The framework produces compact, re-issuable templates and reduces discrimination by the evaluated cross-application similarity score. The evidence establishes neither waveform non-invertibility nor attribute privacy and does not justify secure deployment without further controls. Experimental contexts use deterministic public seeds; secret-key custody and cryptographic key-search resistance remain unmeasured.
Facial Expression Recognition (FER) faces various challenges, including pose variations, occlusions, and diverse illumination conditions. Vision Transformers (ViTs) extract robust features to address these. However, full ViT fine-tuning causes overfitting on small datasets. Moreover, existing ViT adaptation methods typically use only the global class token. This may overlook localized facial micro-movements that are important for accurate expression recognition. In this paper, we propose Symmetric Dual-Head Attention-based Parameter-Efficient Adaptation (SyDHA), an approach to solve this problem. SyDHA explicitly decouples local and global feature processing through spatial attention and global representation branches. The spatial attention branch captures local nuances from patch embeddings, whereas the global representation branch extracts structural context from class tokens. We introduce an inter-classifier symmetry loss to enforce consistent learning across both branches. SyDHA integrates Low-Rank Adaptation (LoRA) and updates only approximately 2.54% of the backbone parameters. An ablation study confirms the effectiveness of each SyDHA component. Component analysis further validates the custom spatial attention mechanism. Sensitivity analysis additionally demonstrates robustness across diverse hyperparameter settings. Visualizations reveal that the spatial attention branch isolates local features, whereas the global representation branch captures a broad context. Feature projections, furthermore, display a distinct semantic clustering of emotions by SyDHA. We also examine the computational efficiency and scalability of SyDHA. We compare SyDHA with 32 existing methods across up to five datasets. It achieves superior performance in 79 out of 81 cases. The code for this work is publicly available at https://github.com/Junelyngdoh/SyDHA.
Surface electromyography (sEMG) has emerged as a promising biometric modality due to its inherent resistance to spoofing and seamless integration with wearable devices. However, existing EMG biometric systems are predominantly task-locked, requiring specific “password” gestures, or restricted to a fixed gesture vocabulary defined during training, conflicting with the unconstrained gesture behavior of realistic wearable deployment. We address this gap by formulating calibration-free cross-gesture EMG biometric authentication: a strict crossges-ture setting where enrollment and authentication gestures can be entirely disjoint, requiring no gesture-specific adaptation at deployment. We propose DADA-Net (Dual Adversarial Dual Alignment Network), which disentangles identity and gesture representations through dual adversarial learning combined with subject- and gesture-conditional distribution alignment. This symmetric design yields gesture-invariant identity representations that generalize to unseen gestures. We evaluate DADA-Net on the GRABMyo and Ninapro DB2 datasets under both identification and verification protocols. Across a comprehensive set of evaluation scenarios, DADA-Net consistently outperforms a range of baselines, with the most pronounced gains on unseen-gesture scenarios. A preliminary cross-day evaluation further shows that, with brief cross-day calibration, DADA-Net recovers most of its within-session performance, indicating its potential for long-term deployment. Comparative analysis of forearm and wrist sEMG further reveals that wrist signals achieve superior authentication performance, supporting their suitability for integration into consumer wearables. This work establishes the feasibility of calibration-free cross-gesture EMG biometric authentication and provides a comprehensive evaluation framework for future research in this paradigm.
Fingerprint biometric systems remain highly vulnerable to presentation attacks, while existing Fingerprint Liveness Detection (FLD) methods suffer from limited cross-sensor and cross-material generalization or rely on computationally expensive adaptation strategies. To address these limitations, we propose FAL-Net, an integrated framework that combines: (i) a ResNet50-based encoder with a hierarchical U-Net decoder enhanced with Soft Spatial Attention (SSA) for spatially selective feature refinement; (ii) a low-rank residual branch integrated into the classification head and jointly trained with the full ResNet50 backbone for domain adaptation; and (iii) a hierarchical multi-scale fusion strategy that combines complementary texture and higher-level representations across network depths. Extensive evaluation on four LivDet benchmarks (2011–2017) demonstrates the effectiveness of the proposed approach. On LivDet 2017, FAL-Net achieves 99.49% average accuracy (ACE = 0.59%) in the initial experimental run. Across three independent repeated runs, it yields mean ACE values of 1.31% ± 0.20% on GreenBit and 1.62% ± 0.35% on Orcanthus, with competitive APCER and BPCER values. In cross-sensor evaluation, FAL-Net further achieves an 8.28-percentage-point improvement in average accuracy over the corresponding baseline across the 12 common sensor-pair directions.
