Large language models (LLMs) have shown impressive general-purpose language capabilities, but their application in specialized domains such as healthcare and law remains limited due to two major challenges, namely, a lack of deep domain-specific knowledge and the inability to incorporate real-time information updates. This paper focuses on addressing these challenges by introducing parameter-sensitive low-rank adaptation (LoRA) and retrieval-augmented generation (RAG), named SensiLoRA-RAG, a two-stage framework designed to enhance LLM performance in domain-specific question-answering tasks. In the first stage, we propose a parameter-sensitive LoRA fine-tuning method that efficiently adapts LLMs to specialized domains using limited high-quality professional data, enabling rapid and resource-efficient specialization. In the second stage, we develop a chain-of-thought RAG mechanism that dynamically retrieves and integrates up-to-date external knowledge, improving the model’s ability to reason with current information and complex domain context. We evaluate our framework on tasks in the medical and legal fields, demonstrating that SensiLoRA-RAG significantly improves answer accuracy, domain relevance, and adaptability compared to baseline methods.

Zero-watermarking is recognized as a pivotal technology for copyright protection. However, the existing audio zero-watermarking schemes are often limited to insufficient robustness and inadequate security. To address these issues, a robust and secure audio zero-watermarking scheme based on novel chaotic map and dual-tree complex wavelet transform (DT-CWT) is proposed. Specifically, a 2-D logistic map with improving chaos complexity (2D-LMICC) is designed. Chaotic performance analysis demonstrates that 2D-LMICC exhibits superior randomness compared to the existing chaotic systems, thus enhancing the security of watermarking system. Then, an audio zero-watermark construction method is proposed. The carrier audio signal is first performed by DT-CWT to obtain coefficients of the low and middle frequency bands with good robustness. After that, these coefficients are processed by singular value decomposition to extract the robust features, and modulated with the watermark to generate plaintext zero-watermark. Finally, the plaintext zero-watermark information is encrypted based on the proposed 2D-LMICC to ensure information security. Extensive experiments demonstrate that the proposed zero-watermarking scheme outperforms the state-of-the-art schemes in resisting desynchronization attacks. It reduces the average bit error rate (BER) by 9.97% and 6.45% under 20% time-scale modification (TSM) and 20% pitch-scale modification (PSM), respectively. Additionally, the ciphertext zero-watermarking exhibits the information entropy near the theoretical maximum of 1 and the adjacent element correlation close to theoretical minimum of 0, effectively ensuring zero-watermarking security.

Traditional audio watermarking schemes do not adequately consider the watermark band misalignment caused by desynchronization attacks, thereby exhibiting vulnerability to such attacks. To address this issue, a novel audio watermarking scheme based on improved quantization index modulation (QIM) and residual-guided scaling factor is proposed. Specifically, the host signal performs discrete cosine transform and singular value decomposition to extract low and medium frequency features with good robustness. Then, the global low-frequency feature information is utilized for quantizer construction, tackling the amplitude scaling-related attacks that classical QIM fails to resist. Next, the chaotic encrypted watermarks are embedded in the mid-frequency features by the improved QIM technique. In the watermark extracting stage, the residual-guided scaling factor is estimated to resist desynchronization attacks. According to the watermark embedding pattern, the misaligned watermark bands caused by the desynchronization attacks are pinpointed by minimizing residuals between the extracted features under different scaling factors and the original watermarked features. After that, the watermark information is directly extracted from these bands. Finally, the performance of the proposed watermarking scheme against desynchronization attacks and common signal processing techniques is theoretically analyzed. Extensive experiments demonstrate that, compared to the state-of-the-art schemes, the proposed scheme provides a more efficient solution to resist desynchronization attacks, while the watermarked audio quality maintains an excellent level.