The Pinching Antenna System (PASS) has recently attracted a lot of attention due to its capability to reconfigure antenna positions along a waveguide dynamically. However, the resulting phase shifts introduce non-convex and highly coupled optimization problems, making closed-form solutions intractable. To overcome this, we propose a Deep Reinforcement Learning (DRL) based Two-Staged Soft Actor-Critic framework (TS-SAC). In the first stage, a Soft Actor-Critic (SAC) agent optimizes the spatial placement of antennas, while the second stage computes the optimal Non-Orthogonal Multiple Access (NOMA) power allocation coefficient using a constrained nonlinear optimization solver. This two-staged approach is evaluated in a two-user downlink NOMA scenario and benchmarked against state-of-theart DRL baselines. Our main aim is to maximize the sum rate of this system. Simulation results show that TS-SAC achieves a 23.7% improvement in average sum rate over Deep Deterministic Policy Gradient (DDPG) at 20 dBm, 11.7% improvement over Twin-Delayed Deep Deterministic Policy Gradient (TD3) and 9% improvement over SAC at the same power level. These results establish the effectiveness and reliability of the proposed approach for next-generation wireless networks.
The proliferation of DeepFake audio technologies has raised significant security and ethical concerns, necessitating robust detection mechanisms. This paper proposes the novel Q-Detect framework that combines a Quantum Vision Transformer (QViT) with a Bidirectional Long Short-Term Memory (BiLSTM) network enhanced by a custom attention mechanism for efficient DeepFake audio detection. The QViT leverages variational quantum circuits within its attention mechanisms to capture intricate spectral patterns from Mel spectrograms, while the BiLSTM network poses temporal dependencies using extracted audio features. Comparative analysis with state-of-theart CNN architectures and conventional Vision Transformers demonstrates the superiority of our proposed approach, highlighting the potential of quantum-enhanced models in audio forensics.
Cross-spectral face recognition aims to match identities across imaging modalities such as NIR (Near-Infrared) -to-VIS (Visible) and Thermal-to-VIS. Large modality differences-loss of texture in NIR and temperature-driven appearance in Thermal-introduce significant domain gaps that degrade performance of VIS-trained systems. This paper provides a systematic architectural review and introduces a Vision Transformerbased model to address these challenges. We categorize heterogeneous face recognition methods into four key paradigms: discriminative embedding learning, generative translation-based approaches, attention and transformer-based models, and hybrid fusion strategies. Inspired by recent architectural evolution, we propose a Vision Transformer with a Domain-Aware Fusion Transformer (DAFT) module that performs cross-spectral multihead attention to preserve identity while aligning modalityspecific representations. Joint optimization of contrastive, triplet, ArcFace, and Maximum Mean Discrepancy (MMD) losses enhances both discriminability and cross-domain consistency. Experiments on Oulu-CASIA and CASIA NIR-VIS 2.0 demonstrate substantial improvements over baselines with +13.73% and +22.36% Rank-1 accuracy gains and 60.4% and 83.1% reductions in domain gap, respectively. This work validates attention-driven spectral fusion as a promising direction for heterogeneous face recognition.
Most Visual Language Models (VLMs) make use of the attention mechanism to achieve consistently high accuracy. However, the quadratic algorithmic complexity (with token length) makes them computationally expensive. CLIPbased models make use of a transformer backbone for both vision and text modalities, thus prioritizing accuracy over inference time. On the other hand, the Vision Mamba (Vim) architecture isolates the attention mechanism used in the transformer and introduces a State Space Model (SSM), which helps in improving the inference time while maintaining accuracy. In this paper, we replace the vision modality of the CLIP model with the Vim model while leaving the text modality untouched and name this model VimCLIP. The proposed VimCLIP, to the best of our knowledge, is the first attempt of a VLM using Vim for cross-modal retrieval and zero-shot classification. The proposed model is trained on the tiny, small, and base variants of the Vim model on the CC3M+CC12M dataset. VimCLIP's performance is assessed on the cross-modal retrieval and its generalization capability on the zero-shot Imagenet dataset. The transformerbased baseline CLIP model outperforms the proposed VimCLIP model on the retrieval task. However, the VimCLIP model consistently beats the transformer-based models on the zero-shot classification task on the Imagenet dataset by approximately 3% on the small variant and about 11% on the tiny variant. This makes the VimCLIP model a reliable and consistent candidate for edge devices. Importantly, VimCLIP's inference time is up to an order of magnitude lower than that of the transformer-based baseline CLIP model. The corresponding codes are available at https://github.com/ambparth/VimCLIP.
In distributed computing with untrusted workers, the assignment of evaluation indices plays a critical role in determining both privacy and robustness. In this work, we study how the placement of unreliable workers within the Numerically Stable Lagrange Coded Computing (NS-LCC) framework influences privacy and the ability to localize Byzantine errors. We derive analytical bounds that quantify how different evaluation-index assignments affect privacy against colluding curious workers and robustness against Byzantine corruption under finite-precision arithmetic. Using these bounds, we formulate optimization problems that identify privacy-optimal and robustness-optimal index placements and show that the resulting assignments are fundamentally different. This exposes that index choices that maximizes privacy degrade error-localization, and vice versa. To jointly navigate this trade-off, we propose a low-complexity greedy assignment strategy that closely approximates the optimal balance between privacy and robustness.
