The prevailing Gaussian Splatting has been adapted for image representation and compression by GaussianImage recently, achieving impressive rendering speeds of 1500–2000 FPS. However, it falls short in rate-distortion performance compared to current state-of-the-art image codecs. We attribute this to its explicit representation and tile-based rasterization process which demands a more efficient utilization of 2D Gaussians for improved performance. In this work, we propose a spectrum-adaptive distribution method to allocate Gaussians in alignment with the complexity of varying image regions. Additionally, we introduce a gradient-based growing strategy for Gaussians to further refine their distribution. Experimental results show that our approach outperforms GaussianImage in image representation quality while maintaining comparable training and rendering speeds. Moreover, by integrating a transform-quantization pipeline for the position attributes of Gaussians, our approach delivers a better rate-distortion curve compared to the INR-based codecs such as COIN and COIN++.

Existing solutions to the joint problem of image deblurring and compression typically adopt a straightforward sequential pipeline, where the two task are handled independently by either compressing the blurred image first or by deblurring before compression. However, such decoupled strategies often suffer from error accumulation and increased computational overhead, leading to suboptimal performance in both compression and restoration. In this work, we propose a joint solution to simultaneously address image deblurring and compression within a learned image compression framework, with the capacity to encode a compact bitstream of the underlying sharp image from its blurred observation. Specifically, our framework has a two-branch encoder architecture, consisting of a main branch and a guidance branch. The guidance branch extracts blur-aware features from the blurry image, which are then fused into the main branch to assist in encoding features that are both bitrate-efficient and essential for accurate sharp restoration. Furthermore, we enhance the fused features in both the spatial and frequency domains. In particular, a JPEG-like transform-quantization mechanism is adopted to modulate the fused features using differentiable quantization matrices in the frequency domain. Extensive experiments on widely-used deblurring datasets demonstrate that our joint solution achieves superior rate-distortion performance, while significantly reducing inference latency compared to existing sequential approaches.

Large language models (LLMs) have demonstrated proficiency in various language tasks but encounter difficulties in specific domain or scenario. These challenges are mitigated through prompt engineering techniques such as retrieval-augmented generation, which improves performance by integrating contextual information. However, concerns regarding the privacy implications of context-aware reasoning architectures persist, particularly regarding the transmission of sensitive data to LLMs service providers, potentially compromising personal privacy. To mitigate these challenges, this paper introduces Tex-tual Differential Privacy, a novel paradigm aimed at safeguarding user privacy in LLMs-based context-aware reasoning. The proposed Differential Embedding Hash algorithm anonymizes sensitive information while maintaining the reasoning capability of LLMs. Additionally, a quantification scheme for privacy loss is proposed to better understand the trade-off between privacy protection and loss. Through rigorous analysis and experimentation, the effectiveness and robustness of the proposed paradigm in mitigating privacy risks associated with context-aware reasoning tasks are demonstrated. This paradigm addresses privacy concerns in context-aware reasoning architectures, enhancing the trust and utility of LLMs in various applications.

Surface electromyography (sEMG) signals find extensive applications in medicine and bioengineering, particularly in rehabilitation and assistive technologies. Gesture classification using sEMG poses challenges such as noise removal, feature extraction, and personalized classification to accommodate individual variations in human physiology. To address these challenges, we propose an sEMG-based gesture classification method by leveraging the nearest centroid classifier and guiding the generation of centroids generated with Soft-DTW as a loss function. Additionally, we apply denoising techniques to the original sEMG signals, including DC offset removal, bandpass filtering, full-wave rectification, and linear envelope extraction. Additionally, we propose a cubic spline for downsampling. With a 1% downsampling rate, our method achieves 89.1% on average and 90.5% at peak and outperforms the state-of-the-art methods.

Online signature verification plays a pivotal role in security infrastructures. However, conventional online signature verification models pose significant risks to data privacy, especially during training processes. To mitigate these concerns, we propose a novel federated learning framework that leverages 1-D Convolutional Neural Networks (CNN) for online signature verification. Furthermore, our experiments demonstrate the effectiveness of our framework regarding 1-D CNN and federated learning. Particularly, the experiment results highlight that our framework 1) minimizes local computational resources; 2) enhances transfer effects with substantial initialization data; 3) presents remarkable scalability. The centralized 1-D CNN model achieves an Equal Error Rate (EER) of 3.33% and an accuracy of 96.25%. Meanwhile, configurations with 2, 5, and 10 agents yield EERs of 5.42%, 5.83%, and 5.63%, along with accuracies of 95.21%, 94.17%, and 94.06%, respectively.

Handwriting signature, as an important behavioral biometric trait, has been broadly adopted for authorization and identity verification. Nowadays, the emergence of consumer-level devices and the development of deep neural networks have vastly facilitated this field. In this paper, we present a practical recurrent neural network-based in-air signature authentication system using smartwatch. The signature is represented by the readings of gyroscope and accelerometer compensated by device attitude readings. The system can extract features from in-air signing motions to verify whether a signature is from an imposter or the genuine user. Moreover, an in-air signature dataset, consisting of 3190 signature pairs from 22 participants, is built for validation. Experimental results demonstrate that our proposed approach achieves an equal error rate(EER) of 0.83%. Besides, we investigate the impact of properties of motion sensory data on in-air signature authentication.