Differential Privacy (DP) has emerged as a rigorous framework for protecting sensitive data in deep learning. However, traditional DP mechanisms often rely on static noise injection strategies that apply uniform perturbations, irrespective of the dynamic privacy-utility trade-offs during training. This one-size-fits-all approach can lead to suboptimal model performance or excessive privacy budget consumption. To address this challenge, we propose a novel adaptive Differential Privacy framework driven by Reinforcement Learning (RL) that dynamically optimizes noise perturbation in deep neural networks. Our framework leverages an RL agent that continuously monitors the training process, including gradient distributions, model convergence, and accumulated privacy loss. Based on these observations, the agent autonomously adjusts the noise magnitude injected into gradients, aiming to optimize the trade-off between privacy leakage and model utility. Specifically, we design a reward function that balances privacy protection (quantified via R & eacute;nyi Differential Privacy (RDP) bounds) and model accuracy, enabling the agent to learn an optimal perturbation policy. Extensive experiments on the CIFAR10 and MNIST benchmarks demonstrate that our method significantly outperforms static DP-SGD and other adaptive baselines. Furthermore, theoretical analysis confirms that the RL-guided adaptation adheres to rigorous (epsilon, delta)-DP guarantees. These results highlight the potential of RL as an adaptive paradigm for privacy-preserving deep learning, offering a flexible and effective alternative to manual parameter tuning.

In the process of minimizing the training loss of machine learning model pursuit,it is very easy to inadvertently remember sensitive private data,which leads to data reconstruction,member reasoning attacks and other security problems.In order to mitigate these risks,differentiated privacy has become a key standard for privacy preserving machine learning.The classical differential privacy depth learning algorithm,differen-tial privacy stochastic gradient descent(DP-SGD),that adapts the standard stochastic gradient descent(SGD)algorithm to incorporate differential privacy,ensuring that the trained model doesn't reveal sensitive infor-mation about individual training data points.However,DP-SGD has the problems of slow convergence speed and large utility loss.We propose an effective solution that is the cooperative combination of selective updat-ing and early vertical gradient disturbance.Selective updating ensures that the training track of the model is aligned with the optimal direction,significantly ac-celerating convergence.Subsequently,the application of vertical gradient perturbation ensures that the model with significantly improved accuracy can be achieved even under strict privacy constraints(small privacy bud-get).Through theoretical analysis and a large number of experiments,this paper proves that DP-VGPSU has superior performance in convergence speed and accu-racy.

Federated Learning (FL) enables collaborative model training without exchanging sensitive local data, ensuring privacy and advancing distributed machine learning. However, in edge scenarios, FL faces challenges of data heterogeneity, device resource constraints, and fairness imbalance among small-data clients, making it difficult to balance performance, efficiency, and fairness. To tackle this, we propose the Discrete Firefly Algorithm (DFA) for fair client selection in FL, mapping clients to fireflies, retaining the brightness attraction"core while adapting to discrete selection. DFA quantifies brightness through data volume and historical contributions, optimizes efficiency with selective sampling, and guarantees fairness for small-data clients. Experiments on MNIST, Fashion-MNIST, and CIFAR-10 demonstrate DFA outperforms baselines: achieving 73.12\((%)\) accuracy on CIFAR-10 (2.56\((%)\) and 10.82\((%)\) higher than random selection and Power-of-Choice), with lower overhead, 2.4\((%)\) performance improvement for small-data clients, and compliance with the principle of contribution-matching benefit.

Support vector machine (SVM) is an efficient classification method in machine learning. The traditional classification model of SVMs may pose a great threat to personal privacy, when sensitive information is included in the training datasets. Principal component analysis (PCA) can project instances into a low-dimensional subspace while capturing the variance of the matrix A as much as possible. There are two common algorithms that PCA uses to perform the principal component analysis, eigenvalue decomposition (EVD) and singular value decomposition (SVD). The main advantage of SVD compared with EVD is that it does not need to compute the matrix of covariance. This study presents a new differentially private SVD algorithm (DPSVD) to prevent the privacy leak of SVM classifiers. The DPSVD generates a set of private singular vectors that the projected instances in the singular subspace can be directly used to train SVM while not disclosing privacy of the original instances. After proving that the DPSVD satisfies differential privacy in theory, several experiments were carried out. The experimental results confirm that our method achieved higher accuracy and better stability on different real datasets, compared with other existing private PCA algorithms used to train SVM.

Support vector machine (SVM) is a robust machine learning method and is widely used in classification. However, the traditional SVM training methods may reveal personal privacy when the training data contains sensitive information. In the training process of SVMs, working set selection is a vital step for the sequential minimal optimization-type decomposition methods. To avoid complex sensitivity analysis and the influence of high-dimensional data on the noise of the existing SVM classifiers with privacy protection, we propose a new differentially private working set selection algorithm (DPWSS) in this paper, which utilizes the exponential mechanism to privately select working sets. We theoretically prove that the proposed algorithm satisfies differential privacy. The extended experiments show that the DPWSS algorithm achieves classification capability almost the same as the original non-privacy SVM under different parameters. The errors of optimized objective value between the two algorithms are nearly less than two, meanwhile, the DPWSS algorithm has a higher execution efficiency than the original non-privacy SVM by comparing iterations on different datasets. To the best of our knowledge, DPWSS is the first private working set selection algorithm based on differential privacy.

Support vector machines (SVMs) are among the most robust and accurate methods in all well-known machine learning algorithms, especially for classification. The SVMs train a classification model by solving an optimization problem to decide which instances in the training datasets are the support vectors (SVs). However, SVs are intact instances taken from the training datasets and directly releasing the classification model of the SVMs will carry significant risk to the privacy of individuals, when the training datasets contain sensitive information. In this paper, we study the problem of how to release the classification model of kernel SVMs while preventing privacy leakage of the SVs and satisfying the requirement of privacy protection. We propose a new differentially private algorithm for the kernel SVMs based on the exponential and Laplace hybrid mechanism named DPKSVMEL. The DPKSVMEL algorithm has two major advantages compared with existing private SVM algorithms. One is that it protects the privacy of the SVs by postprocessing and the training process of the non-private kernel SVMs does not change. Another is that the scoring function values are directly derived from the symmetric kernel matrix generated during the training process and does not require additional storage space and complex sensitivity analysis. In the DPKSVMEL algorithm, we define a similarity parameter to denote the correlation or distance between the non-SVs and every SV. And then, every non-SV is divided into a group with one of the SVs according to the maximal value of the similarity. Under some certain similarity parameter value, we replace every SV with a mean value of the top-k randomly selected most similar non-SVs within the group by the exponential mechanism if the number of non-SVs is greater than k. Otherwise, we add random noise to the SVs by the Laplace mechanism. We theoretically prove that the DPKSVMEL algorithm satisfies differential privacy. The extensive experiments show the effectiveness of the DPKSVMEL algorithm for kernel SVMs on real datasets; meanwhile, it achieves higher classification accuracy than existing private SVM algorithms.

Aiming to provide more information about the behaviors between groups or patterns between clusters in social networks, we propose a two-step differentially private method to release the distribution of clustering coefficients across communities. The DPLM algorithm improves a Louvain method to partition one network using an exponential mechanism. We introduce an absolute gain of modularity to sanitize neighboring communities. Otherwise, the algorithm is difficult to converge due to the randomness introduced. The DPCC algorithm charts the noisy distribution of clustering coefficients as a histogram, which presents the results in an intuitive manner. We conduct experiments on three real-world datasets to evaluate the proposed method. The experimental results indicate that the proposed method provides valuable distribution results while guaranteeing ε-differential privacy. Moreover, the DPLM algorithm can obtain better modularity for the networks.

Focusing on the privacy issues in recommender systems, we propose a framework containing two perturbation methods for differentially private collaborative filtering to prevent the threat of inference attacks against users. To conceal individual ratings and provide valuable predictions, we consider some representative algorithms to calculate the predicted scores and provide specific solutions for adding Laplace noise. The DPI (Differentially Private Input) method perturbs the original ratings, which can be followed by any recommendation algorithms. By contrast, the DPM (Differentially Private Manner) method is based on the original ratings, which perturbs the measurements during implementation of the algorithms and releases the predicted scores. The experimental results showed that both methods can provide valuable prediction results while guaranteeing DP, which suggests it is a feasible solution and can be competent to make private recommendations.

To publish social graphs with differential privacy guarantees for reproducing valuable results of scientific researches, we study a workflow for graph synthesis and propose an improved approach based on weighted Privacy integrated query (wPINQ). The workflow starts with a seed graph to fit the noisy degree sequence, which essentially is the 1K-graph. In view of the inaccurate assortativity coefficient, we truncate the workflow to replace the seed graph with an optimal one by doing target 1K-rewiring while preserving the 1K-distribution. Subsequently, Markov chain Monte Carlo employs the new seed graph as the initial state, and proceeds step by step guided by the information of Triangles by intersect to increase the number of triangles in the synthetic graphs. The experimental results show that the proposed algorithm achieves better performance for the published social graphs.

Social networks can be analyzed to discover important social issues; however, it will cause privacy disclosure in the process. The edge weights play an important role in social graphs, which are associated with sensitive information (e.g., the price of commercial trade). In the paper, we propose the MB-CI (Merging Barrels and Consistency Inference) strategy to protect weighted social graphs. By viewing the edge-weight sequence as an unattributed histogram, differential privacy for edge weights can be implemented based on the histogram. Considering that some edges have the same weight in a social network, we merge the barrels with the same count into one group to reduce the noise required. Moreover, k-indistinguishability between groups is proposed to fulfill differential privacy not to be violated, because simple merging operation may disclose some information by the magnitude of noise itself. For keeping most of the shortest paths unchanged, we do consistency inference according to original order of the sequence as an important postprocessing step. Experimental results show that the proposed approach effectively improved the accuracy and utility of the released data.

In view of the intrusion detection system (IDS)'s problem of the large number of alerts and the high false positive rate,an improved model of Snort-based intrusion detection alerts analysis system was designed. It bases on Snort,an open source IDS,and establishes normal and abnormal alerts behavior patterns. So the detection efficiency and unknown attacks detection capability of IDS are raised to some extent.

Against the high NOx emission of a thermal power plant using low volatile coal,it was applied to reduce the NOx emission through low nitrogen boiler alteration and SOFA vertical depth grading technology on the whole furnace.Through the analysis of the influence of over-fire air flow,furnace oxygen amount,the way to powder,the operation of coal pulverized,the result shows that the NOx emission can be decreased to 53.9%,however,too large over-fire air flow volume and too low furnace oxygen amount will decrease boiler operating efficiency.

With the unceasing expansion of postgraduate enrollment scale and the continuous progress of educational reform,the standardizing demand of postgraduate training administration course is higher and higher,our university has established postgraduate management information systems based on Web in order to standardize the whole course of postgraduate educational training,the postgraduate management information systems plays an very important role in enhancing work level and work efficiency of postgraduate educational administration and postgraduate training quality.

In view of the intrusion detection system(IDS)'s characteristic,it analyzed the current research situation of the data mining(DM) application in intrusion detection(ID),and making use of superiority of DM in disposing massive alerts data,designed a model of intrusion alerts analysis system.It reduces the number of alerts,raises the detection efficiency and practicability by analyzing intrusion alerts.

In view of the current development of adaptive Web site,an executive framework of HMM-based adaptive Web site is put forward.Through semantic analysis of user access sequences,latent information requirement contained in paths is extracted.This model adjusts Web site,recommends and pre-fetches future requests accordingly.It upgrades the adaptive performance of Web site.

In allusion to the efficiency problem of the actual association rules mining algorithms in the process of disposing massive data,an improved Recursive Elimination based on linked List(LRE) based on one-dimensional linked list was put forward in this paper.A model of intrusion alerts analysis system with LRE algorithm was designed.Finally,the validity in the aspect of reducing the number of alerts and the false positive rate has been showed by the experiments.

This paper puts forward a model of Snort intrusion detection system based on data mining technology. It takes the frame of Snort intrusion detection system as the foundation and makes use of data mining technology to construct the normal behavior pattern of system, which filters a majority of normal data in the network data stream, therefore it improves the detecting efficiency of Snort. At the same time, it also can detect unknown attack to some extent, so it enhances the detecting ability of Snort to unknown attack. These tests showed this model is valid.