Shift-Accumulate Attention: Multiplier-Free Query--Key Products for Transformer Decoding
研究提出了一种无乘法器的Shift-Accumulate Attention方法,通过将键缓存量化为有符号的2的幂固定点编码,使QK^T中的每个标量乘法变为符号翻转、位移和整数累加,从而提高Transformer解码效率。
研究提出了一种无乘法器的Shift-Accumulate Attention方法,通过将键缓存量化为有符号的2的幂固定点编码,使QK^T中的每个标量乘法变为符号翻转、位移和整数累加,从而提高Transformer解码效率。
研究通过Phoenix识别器和Athar审查流程解决阿拉伯手稿转录问题,包括应对不同书写风格、布局变化及保留不确定性读取。
This work addresses the challenges of low reconstruction accuracy and poor generalization in 3D organ modeling from macroscopic slice imaging, which arise due to data scarcity and large deformations. To overcome these limitations, the authors propose a two-stage hybrid registration framework: an initial global rigid alignment is achieved through Optimal Slice Matching (OCM) combined with Hough transform, followed by local non-rigid deformation estimation using explicit geometric priors integrated into a lightweight, modified VoxelMorph network. By hierarchically decoupling global optimization from local refinement, the method significantly outperforms single-stage baselines—even when trained on only 40 kidney specimens—yielding more accurate, anatomically plausible, and reproducible multimodal 3D reconstructions suitable for surgical planning and medical education.
This work proposes a novel image encryption framework that integrates fractal geometry with Fourier transform, addressing the longstanding challenge of simultaneously achieving high security, image fidelity, and computational efficiency in traditional methods. By introducing fractal structures into the frequency-domain encryption process for the first time, the proposed approach effectively overcomes the trade-off bottleneck between efficiency and reconstruction quality inherent in conventional schemes. Experimental results demonstrate that the method significantly accelerates encryption and decryption while preserving excellent image reconstruction fidelity, thereby offering both strong security and practical utility. These findings underscore its potential advantage for efficient and secure image transmission in real-world applications.
Traditional CAPTCHAs struggle to balance security and usability. This paper proposes a novel hybrid CAPTCHA system integrating generative AI with keystroke dynamics: large language models (LLMs) dynamically generate semantic cognition challenges, while users’ keystroke timing features are simultaneously captured and analyzed to establish a dual-modal “cognitive-behavioral” discrimination mechanism. To our knowledge, this is the first work unifying LLM-driven dynamic semantic verification with biometric-level input rhythm analysis, effectively thwarting paste attacks, scripted automation, and end-to-end AI-based bypasses. Experimental results show that the system achieves a 92.3% human success rate—indicating excellent usability—while maintaining a 99.1% bot detection accuracy, significantly outperforming state-of-the-art text- and image-based CAPTCHAs. This work establishes a scalable, adaptively robust multimodal security paradigm for next-generation human-bot differentiation.
研究提出了一种无乘法器的Shift-Accumulate Attention方法,通过将键缓存量化为有符号的2的幂固定点编码,使QK^T中的每个标量乘法变为符号翻转、位移和整数累加,从而提高Transformer解码效率。
研究通过Phoenix识别器和Athar审查流程解决阿拉伯手稿转录问题,包括应对不同书写风格、布局变化及保留不确定性读取。
This work addresses the challenges of low reconstruction accuracy and poor generalization in 3D organ modeling from macroscopic slice imaging, which arise due to data scarcity and large deformations. To overcome these limitations, the authors propose a two-stage hybrid registration framework: an initial global rigid alignment is achieved through Optimal Slice Matching (OCM) combined with Hough transform, followed by local non-rigid deformation estimation using explicit geometric priors integrated into a lightweight, modified VoxelMorph network. By hierarchically decoupling global optimization from local refinement, the method significantly outperforms single-stage baselines—even when trained on only 40 kidney specimens—yielding more accurate, anatomically plausible, and reproducible multimodal 3D reconstructions suitable for surgical planning and medical education.
This work proposes a novel image encryption framework that integrates fractal geometry with Fourier transform, addressing the longstanding challenge of simultaneously achieving high security, image fidelity, and computational efficiency in traditional methods. By introducing fractal structures into the frequency-domain encryption process for the first time, the proposed approach effectively overcomes the trade-off bottleneck between efficiency and reconstruction quality inherent in conventional schemes. Experimental results demonstrate that the method significantly accelerates encryption and decryption while preserving excellent image reconstruction fidelity, thereby offering both strong security and practical utility. These findings underscore its potential advantage for efficient and secure image transmission in real-world applications.
Traditional CAPTCHAs struggle to balance security and usability. This paper proposes a novel hybrid CAPTCHA system integrating generative AI with keystroke dynamics: large language models (LLMs) dynamically generate semantic cognition challenges, while users’ keystroke timing features are simultaneously captured and analyzed to establish a dual-modal “cognitive-behavioral” discrimination mechanism. To our knowledge, this is the first work unifying LLM-driven dynamic semantic verification with biometric-level input rhythm analysis, effectively thwarting paste attacks, scripted automation, and end-to-end AI-based bypasses. Experimental results show that the system achieves a 92.3% human success rate—indicating excellent usability—while maintaining a 99.1% bot detection accuracy, significantly outperforming state-of-the-art text- and image-based CAPTCHAs. This work establishes a scalable, adaptively robust multimodal security paradigm for next-generation human-bot differentiation.