Fine-tuning large language models on downstream tasks is crucial for realizing their cross-domain potential but often relies on sensitive data, raising privacy concerns. Differential privacy (DP) offers rigorous privacy guarantees and has been widely adopted in fine-tuning; however, naively injecting noise across the high-dimensional parameter space creates perturbations with large norms, degrading performance and destabilizing training. To address this issue, we propose DP-SFT, a two-stage subspace fine-tuning method that substantially reduces noise magnitude while preserving formal DP guarantees. Our intuition is that, during fine-tuning, significant parameter updates lie within a low-dimensional, task-specific subspace, while other directions change minimally. Hence, we only inject DP noise into this subspace to protect privacy without perturbing irrelevant parameters. In phase one, we identify the subspace by analyzing principal gradient directions to capture task-specific update signals. In phase two, we project full gradients onto this subspace, add DP noise, and map the perturbed gradients back to the original parameter space for model updates, markedly lowering noise impact. Experiments on multiple datasets demonstrate that DP-SFT enhances accuracy and stability under rigorous DP constraints, accelerates convergence, and achieves substantial gains over DP fine-tuning baselines.

Abstract In current research on reversible visible watermarking algorithm, the original visible watermark image plays an important auxiliary role, and some algorithms also entirely depend on it to restore host image without any distortion. Therefore, in order to realize semi-blind reversible visible watermarking algorithm, the conventional reversible watermarking algorithm is used to embed compressed visible watermark image data into non-visible-watermarked region of host image. However, the amount of compressed image data obtained by conventional image compression algorithm is relatively large. Therefore, a method based on vectorization compression for the visible watermark image is proposed in this paper. Firstly, it performs edge detection on visible watermark image to obtain a discrete points set $\Gamma $ of vector contour curve. Then, the discrete points in $\Gamma $ are simplified by improved Douglas–Peucker algorithm, after that it obtains compressed vector contour data of visible watermark image. In addition, a reversible visible watermarking algorithm based on convolutional relief and image alpha fusion is proposed, which realizes reversible embedding of visible watermark image and lossless restoration of host image. The experimental results show that the proposed vectorization compression method has more advantages than traditional image compression algorithms, which greatly reduces the storage space of visible watermark image with high fidelity. Additionally, the embedded watermarking image has translucent 3D relief effect, and the fusion of host image and visible watermark image becomes more natural and harmonious.

With the transmission and sharing of data in the Internet of Things (IoT), while bringing development to life and the economy, it also inevitably threatens the data copyright protection and authentication. Digital watermarking technology can provide an effective solution for copyright protection by embedding the watermark in the data to prove the copyright attribution. The existing methods of digital watermarking in IoT mainly target multimedia, without considering the copyright authentication in database data. Unlike multimedia information, the database does not focus on the subjective visual perception when using the data, but rather on the potential values unlocked from the data through algorithms such as data mining. Therefore, we propose a new database watermarking algorithm based on decision tree shift correction (DTSC), considering the data copyright authentication and usability when applying for data mining algorithm. The algorithm adjusts the watermarked data by the DTSC method and makes the watermarked decision tree identical to the original in the iteration process. It solves the problem of database data copyright authentication in IoT and ensures the usability of the data when used for decision tree model construction. From the simulation results, it can be seen that the proposed method ensures the usability of the data for the classification and regression tree decision tree algorithm while embedding the watermark in the database data, and the data distortion of the proposed method does not differ from that of the traditional watermarking algorithm.

Aiming at the problem that the robustness, visibility, and transparency of the existing visible watermarking technologies are difficult to achieve a balance, this paper proposes an adaptive embedding method for visible watermarking. Firstly, the salient region of the host image is detected based on superpixel detection. Secondly, the flat region with relatively low complexity is selected as the embedding region in the nonsalient region of the host image. Then, the watermarking strength is adaptively calculated by considering the gray distribution and image texture complexity of the embedding region. Finally, the visible watermark image is adaptively embedded into the host image with slight adjustment by just noticeable difference (JND) coefficient. The experimental results show that our proposed method improves the robustness of visible watermarking technology and greatly reduces the risk of malicious removal of visible watermark image. Meanwhile, a good balance between the visibility and transparency of the visible watermark image is achieved, which has the advantages of high security and ideal visual effect.

Robustness is the most important factor that limits the practical application of reversible watermarking. To deal with this issue, several robust reversible watermarking (RRW) techniques have been proposed. Among them, the two-stage RRW framework proposed by Coltuc et al. is a promising one. In the first state of this framework, a robust watermark is embedded into the cover image to provide robustness, and then in the second stage, the information enabling revert the robust embedding is reversibly embedded into the already marked image to guarantee the reversibility. However, because of using the same area for these two embedding stages, the robustness in the first stage is seriously weakened by the reversible embedding. As a result, this elegant method is not effective as expected. Based on this consideration, this paper proposes an independent embedding domain (ED)-based two-stage RRW. The cover image is first transformed into two independent EDs, and then the robust and reversible watermarks are embedded into each domain separately. The carrier derived from the first embedding stage that carrying the robust watermark will not change after the reversible embedding, and thus, the robustness of the first stage is well preserved. By the proposed method, the embedding performance of the original two-stage RRW is significantly enhanced. Moreover, the proposed method is experimentally verified better than some other state-of-the-art RRW methods.

Privacy protection has recently attracted increasingly attention in location-based services (LBS) with the development of wireless communication and mobile positioning technology. Many location-based cloaking algorithms are proposed to protect location privacy of mobile users. However, there is a scenario where the anonymous server only utilizes users defined as current users in our algorithm, who are requiring LBS requests at the current time, to form the cloaking region. If there are not enough current users at the current time, the generated cloaking region will be larger than expected. To solve this problem, we propose the probability-base location prediction ( PLP ) algorithm. Mobile users are divided into current users and historical users. According to predicting probability, historical users are also utilized to generate the cloaking region when current users are not enough to achieve the k-anonymity model. Experimental results show that PLP can predict the cloaking region as high as 90% and the cloaking region can also decrease remarkably.

As one of the most efficient multi-secret visual cryptography (MVC) schemes, the tagged visual cryptography (TVC) is capable of hiding tag images into randomly selected shares. However, the encoding processes of TVC and other MVC schemes bring distortion to shares, which definitely lowers the visual quality of the decoded secret image. This letter proposes an extended TVC scheme, named as lossless TVC (LTVC). Specifically, “lossless” means that the proposed LTVC scheme encodes the tag image without affecting the rebuilt secret image, i.e., the decoded secret image of LTVC has the same visual quality with that of the conventional VC scheme . Moreover, we propose the probabilistic LTVC (P-LTVC) to solve the potential security problem of LTVC. Finally, the superiority of the proposed scheme is experimentally verified.

In this letter, an efficient integer transform based reversible watermarking is proposed. We first show that Tian's difference expansion (DE) technique can be reformulated as an integer transform. Then, a generalized integer transform and a payload-dependent location map are constructed to extend the DE technique to the pixel blocks of arbitrary length. Meanwhile, the distortion can be controlled by preferentially selecting embeddable blocks that introduce less distortion. Finally, the superiority of the proposed method is experimental verified by comparing with other existing schemes.