Whispered Tuning: Data Privacy Preservation in Fine-Tuning LLMs through Differential Privacy

The proliferation of Large Language Models (LLMs) across various sectors underscored the urgency of addressing potential privacy breaches. Vulnerabilities, such as prompt injection attacks and other adversarial tactics, could make these models inadvertently disclose their training data. Such disclosures could compromise personal identifiable information, posing significant privacy risks. In this paper, we proposed a novel multi-faceted approach called Whispered Tuning to address privacy leaks in large language models (LLMs). We integrated a PII redaction model, differential privacy techniques, and an output filter into the LLM fine-tuning process to enhance confidentiality. Additionally, we introduced novel ideas like the Epsilon Dial for adjustable privacy budgeting for differentiated Training Phases per data handler role. Through empirical validation, including attacks on non-private models, we demonstrated the robustness of our proposed solution SecureNLP in safeguarding privacy without compromising utility. This pioneering methodology significantly fortified LLMs against privacy infringements, enabling responsible adoption across sectors.

Higher-order differential variational principle and differential equations of motion for mechanical systems in event space

In this paper we study the higher-order differential variational principle and differential equations of motion for mechanical systems in event space. Based on the higher-order d＇Alembert principle of the system, the higher-order velocity energy and the higher-order acceleration energy of the system in event space are defined, the higher-order d＇Alembert- Lagrange principle of the system in event space is established, and the parametric forms of Euler-Lagrange, Nielsen and Appell for this principle are given. Finally, the higher-order differential equations of motion for holonomic systems in event space are obtained.

Differential Collision Attack on Reduced FOX Block Cipher

This paper presents a method for differen- tial collision attack of reduced FOX block cipher based on 4-round distinguishing property. It can be used to attack 5, 6 and 7-round FOX64 and 5-round FOX128. Our attack has a precomputation phase, but it can be obtained before attack and computed once for all. This attack on the reduced to 4-round FOX64 requires only 7 chosen plaintexts, and performs 242.8 4-round FOX64 encryptions. It could be extended to 5 （6, 7）-round FOX64 by a key exhaustive search behind the fourth round.

Complements to the Theory of Higher-Order Types of Asymptotic Variation for Differentiable Functions

The purpose of this paper is to add some complements to the general theory of higher-order types of asymptotic variation developed in two previous papers so as to complete our elementary (but not too much!) theory in view of applications to the theory of finite asymptotic expansions in the real domain, the asymptotic study of ordinary differential equations and the like. The main results concern: 1) a detailed study of the types of asymptotic variation of an infinite series so extending the results known for the sole power series;2) the type of asymptotic variation of a Wronskian completing the many already-published results on the asymptotic behaviors of Wronskians;3) a comparison between the two main standard approaches to the concept of “type of asymptotic variation”: via an asymptotic differential equation or an asymptotic functional equation;4) a discussion about the simple concept of logarithmic variation making explicit and completing the results which, in the literature, are hidden in a quite-complicated general theory.

Recurrent Transient Ischemic Attacks Revealing Cerebral Amyloid Angiopathy: A Comprehensive Case

This case report investigates the manifestation of cerebral amyloid angiopathy (CAA) through recurrent Transient Ischemic Attacks (TIAs) in an 82-year-old patient. Despite initial diagnostic complexities, cerebral angiography-MRI revealed features indicative of CAA. Symptomatic treatment resulted in improvement, but the patient later developed a fatal hematoma. The discussion navigates the intricate therapeutic landscape of repetitive TIAs in the elderly with cardiovascular risk factors, emphasizing the pivotal role of cerebral MRI and meticulous bleeding risk management. The conclusion stresses the importance of incorporating SWI sequences, specifically when suspecting a cardioembolic TIA, as a diagnostic measure to explore and exclude CAA in the differential diagnosis. This case report provides valuable insights into these challenges, highlighting the need to consider CAA in relevant cases.

FMHADP:Design of an Efficient Pre-Forensic Layer for Mitigating Hybrid Attacks via Deep Learning Pattern Analysis

Network attack detection and mitigation require packet collection,pre-processing,feature analysis,classification,and post-processing.Models for these tasks sometimes become complex or inefficient when applied to real-time data samples.To mitigate hybrid assaults,this study designs an efficient forensic layer employing deep learning pattern analysis and multidomain feature extraction.In this paper,we provide a novel multidomain feature extraction method using Fourier,Z,Laplace,Discrete Cosine Transform(DCT),1D Haar Wavelet,Gabor,and Convolutional Operations.Evolutionary method dragon fly optimisation reduces feature dimensionality and improves feature selection accuracy.The selected features are fed into VGGNet and GoogLeNet models using binary cascaded neural networks to analyse network traffic patterns,detect anomalies,and warn network administrators.The suggested model tackles the inadequacies of existing approaches to hybrid threats,which are growing more common and challenge conventional security measures.Our model integrates multidomain feature extraction,deep learning pattern analysis,and the forensic layer to improve intrusion detection and prevention systems.In diverse attack scenarios,our technique has 3.5% higher accuracy,4.3% higher precision,8.5% higher recall,and 2.9% lower delay than previous models.

Typical adaptive neural control for hypersonic vehicle based on higher-order filters

A typical adaptive neural control methodology is used for the rigid body model of the hypersonic vehicle. The rigid body model is divided into the altitude subsystem and the velocity subsystem. The proportional integral differential(PID) controller is introduced to control the velocity track. The backstepping design is applied for constructing the controllers for the altitude subsystem.To avoid the explosion of differentiation from backstepping, the higher-order filter dynamic is used for replacing the virtual controller in the backstepping design steps. In the design procedure,the radial basis function(RBF) neural network is investigated to approximate the unknown nonlinear functions in the system dynamic of the hypersonic vehicle. The simulations show the effectiveness of the design method.

Differential Quadrature Method for Bending Problem of Plates with Transverse Shear Effects

A differential quadrature (DQ) method for orthotropic plates was proposed based on Reddy' s theory of plates with the effects of the higher-order transverse shear deformations. Wang-Bert's DQ approach was also further extended to handle the boundary conditions of plates. The computational convergence was studied, and the numerical results were obtained for different grid spacings and compared with the existing results. The results show that the DQ method is fairly reliable and effective.

Privacy Protection Algorithm for the Internet of Vehicles Based on Local Differential Privacy and Game Model

In recent years,with the continuous advancement of the intelligent process of the Internet of Vehicles(IoV),the problem of privacy leakage in IoV has become increasingly prominent.The research on the privacy protection of the IoV has become the focus of the society.This paper analyzes the advantages and disadvantages of the existing location privacy protection system structure and algorithms,proposes a privacy protection system structure based on untrusted data collection server,and designs a vehicle location acquisition algorithm based on a local differential privacy and game model.The algorithm first meshes the road network space.Then,the dynamic game model is introduced into the game user location privacy protection model and the attacker location semantic inference model,thereby minimizing the possibility of exposing the regional semantic privacy of the k-location set while maximizing the availability of the service.On this basis,a statistical method is designed,which satisfies the local differential privacy of k-location sets and obtains unbiased estimation of traffic density in different regions.Finally,this paper verifies the algorithm based on the data set of mobile vehicles in Shanghai.The experimental results show that the algorithm can guarantee the user’s location privacy and location semantic privacy while satisfying the service quality requirements,and provide better privacy protection and service for the users of the IoV.

Fourth-Order Splitting Methods for Time-Dependant Differential Equations

This study was suggested by previous work on the simulation of evolution equations with scale-dependent processes,e.g.,wave-propagation or heat-transfer,that are modeled by wave equations or heat equations.Here,we study both parabolic and hyperbolic equations.We focus on ADI (alternating direction implicit) methods and LOD (locally one-dimensional) methods,which are standard splitting methods of lower order,e.g.second-order.Our aim is to develop higher-order ADI methods,which are performed by Richardson extrapolation,Crank-Nicolson methods and higher-order LOD methods,based on locally higher-order methods.We discuss the new theoretical results of the stability and consistency of the ADI methods.The main idea is to apply a higher- order time discretization and combine it with the ADI methods.We also discuss the dis- cretization and splitting methods for first-order and second-order evolution equations. The stability analysis is given for the ADI method for first-order time derivatives and for the LOD (locally one-dimensional) methods for second-order time derivatives.The higher-order methods are unconditionally stable.Some numerical experiments verify our results.

DIFFERENTIAL QUADRATURE METHOD FOR BENDING OF ORTHOTROPIC PLATES WITH FINITE DEFORMATION AND TRANSVERSE SHEAR EFFECTS

Based on the Reddy's theory of plates with the effect of higher-order shear deformations, the governing equations for bending of orthotropic plates with finite deformations were established. The differential quadrature (DQ) method of nonlinear analysis to the problem was presented. New DQ approach, presented by Wang and Bert (DQWB), is extended to handle the multiple boundary conditions of plates. The techniques were also further extended to simplify nonlinear computations. The numerical convergence and comparison of solutions were studied. The results show that the DQ method presented is very reliable and valid. Moreover, the influences of geometric and material parameters as well as the transverse shear deformations on nonlinear bending were investigated. Numerical results show the influence of the shear deformation on the static bending of orthotropic moderately thick plate is significant.

Influence of multi-photon pulses on practical differential-phase-shift quantum key distribution

The influence of multi-photon pulses on practical differcBtial-phase-shift quantum key distribution （DPS-Qt（D） is analysed. We have estimated the information which Eve obtained by PNS （photon-number splitting） attack and BS （beam splitting） attack. The result indicates that the PNS attack and BS attack will not limit the transmission distance as long as we select an appropriate mean photon number. Also, the maximum mean photon number under BS attack in practical DPS-QKD system and the set of practical assumptions about Eve＇s capabilities are presented.

Parallel Evolutionary Modeling for Nonlinear Ordinary Differential Equations

We introduce a new parallel evolutionary algorithm in modeling dynamic systems by nonlinear higher-order ordinary differential equations (NHODEs). The NHODEs models are much more universal than the traditional linear models. In order to accelerate the modeling process, we propose and realize a parallel evolutionary algorithm using distributed CORBA object on the heterogeneous networking. Some numerical experiments show that the new algorithm is feasible and efficient.

Round-robin differential quadrature phase-shift quantum key distribution

Recently,a round-robin differential phase-shift（RRDPS） protocol was proposed[Nature 509,475（2014）],in which the amount of leakage is bounded without monitoring the signal disturbance.Introducing states of the phase-encoded Bennett-Brassard 1984 protocol（PE-BB84） to the RRDPS,this paper presents another quantum key distribution protocol called round-robin differential quadrature phase-shift（RRDQPS） quantum key distribution.Regarding a train of many pulses as a single packet,the sender modulates the phase of each pulse by one of {0,π/2,π,3π/2},then the receiver measures each packet with a Mach-Zehnder interferometer having a phase basis of 0 or π/2.The RRDQPS protocol can be implemented with essential similar hardware to the PE-BB84,so it has great compatibility with the current quantum system.Here we analyze the security of the RRDQPS protocol against the intercept-resend attack and the beam-splitting attack.Results show that the proposed protocol inherits the advantages arising from the simplicity of the RRDPS protocol and is more robust against these attacks than the original protocol.

SKINNY的差分故障攻击与ForkAE的密钥恢复攻击

作为LWC竞赛的候选算法之一,ForkAE是基于叉形密码结构的一系列轻量级认证加密算法,其中使用的加密原语为轻量级可调分组密码族SKINNY.本文首先给出了一种对SKINNY族内各算法进行差分故障攻击的方法.对于SKINNY-64-64和SKINNY-64-128,在算法倒数第三轮注入随机半字节故障,理论上平均通过2.32次随机半字节故障注入即可得知连续两轮共4个半字节的信息.通过对多个位置的信息获取,理论上平均通过9.23次随机半字节故障注入即可恢复单轮的64 bit轮密钥,结合密钥扩展算法即可恢复全部64 bit的主密钥.利用类似的方法攻击SKINNY-128-128和SKINNY-128-256,在算法倒数第三轮注入随机字节故障,理论上平均通过2.4次随机字节故障注入即可得知连续两轮共4个半字节的信息,平均通过9.56次随机字节故障注入即可恢复128 bit的主密钥.攻击SKINNY-64-192或SKINNY-128-384时需要额外对倒数第五轮进行攻击,分别需要18.52次随机半字节故障注入和19.18次随机字节故障注入.在对SKINNY完成的差分故障分析的基础上,本文给出了一种对ForkAE进行密钥恢复的方法,理论上仅需要1次对故障加密机的询问即可恢复ForkAE的主密钥.

基于本地差分隐私的异步横向联邦安全梯度聚合方案

联邦学习作为一种新兴的分布式机器学习框架,通过在用户私有数据不出域的情况下进行联合建模训练,有效地解决了传统机器学习中的数据孤岛和隐私泄露问题。然而,联邦学习存在着训练滞后的客户端拖累全局训练速度的问题,异步联邦学习允许用户在本地完成模型更新后立即上传到服务端并参与到聚合任务中,而无需等待其他用户训练完成。然而,异步联邦学习也存在着无法识别恶意用户上传的错误模型,以及泄露用户隐私的问题。针对这些问题,该文设计一种面向隐私保护的异步联邦的安全梯度聚合方案(SAFL)。用户采用本地差分隐私策略,对本地训练的模型添加扰动并上传到服务端,服务端通过投毒检测算法剔除恶意用户,以实现安全聚合(SA)。最后,理论分析和实验表明在异步联邦学习的场景下,提出的方案能够有效识别出恶意用户,保护用户的本地模型隐私,减少隐私泄露风险,并相对于其他方案在模型的准确率上有较大的提升。

基于博弈论的弹目攻防决策方法研究

针对空战环境中弹目攻防双方的对抗特性,提出了一种基于博弈论的弹目攻防决策方法。基于导弹目标运动数学关系得到状态方程,根据弹目攻防对抗机理建立“一对一导弹-目标”动态博弈模型,确定弹目双方策略集与收益矩阵,提出混合策略纳什均衡求解方法,并结合模型滚动预测方法获得该策略空间的纳什均衡点。算例仿真结果表明,基于混合策略下导弹制导律为该策略空间的纳什均衡点,且该方法可以减小导弹对目标的脱靶量,提高导弹的命中精度,为导弹攻防作战提供了依据。

DoS攻击下抛物型多智能体系统的输出反馈领导-跟随一致性

为了解决抛物型多智能体系统(MASs)在拒绝服务(DoS)攻击下的一致性问题,提出了一种基于输出反馈方法的MASs领导-跟随一致性控制协议来应对恶意攻击者发起的DoS攻击,从而实现智能体在无向通信拓扑图上的领导-跟随一致性。对DoS攻击的频率和持续时间进行了分析,利用Lyapunov稳定性定理推导了抛物型偏微分系统的稳定条件,证明了触发时间间隔不为零,避免了芝诺行为的发生,并通过仿真算例验证了理论结果的有效性。

基于差分隐私与模型聚类的安全联邦学习方案

联邦学习中的模型安全以及客户隐私是亟待解决的重要挑战。为了同时应对这2大挑战,提出了一项基于差分隐私与模型聚类的联邦学习方案,该方案兼顾模型安全与隐私保护。通过在客户更新中引入局部差分隐私扰乱客户上传的参数以保护客户的隐私数据。为保证对加噪模型更新的精准聚类,首次定义余弦梯度作为聚类指标,并根据聚类结果精准定位恶意模型。最后引入全局差分隐私以抵御潜在的后门攻击。通过理论分析得到全局噪声的噪声边界,并证明了本方案引入的噪声总量低于经典模型安全方案所引入的噪声总量。实验结果表明,本方案能够达成在精度、鲁棒以及隐私3方面的预期目标。

RAIN-128算法的中间相遇攻击

RAIN是一族SPN结构的轻量级分组密码算法,该算法具有软硬件实现效率高、安_全性强等特点。中间相遇攻击被广泛应用于分组密码算法的安全性分析中。该文通过分析RAIN-128的结构特性和截断差分特征,利用差分枚举技术分别构造了4轮和6轮中间相遇区分器,给出了8轮及10轮的中间相遇攻击。当攻击轮数为8轮时,预计算阶段的时间复杂度为2^(68)次8轮RAIN-128加密,存储复杂度为2^(75)bit,在线攻击阶段的时间复杂度为2^(109)次8轮加密,数据复杂度是2^(72)个选择明文;当攻击轮数为10轮时,预计算阶段的时间复杂度为2^(214)次10轮加密,存储复杂度为2^(219)bit,在线攻击阶段的时间复杂度为2^(109)次10轮加密,数据复杂度是2^(72)个选择明文,分析结果显示,RAIN-128可以抵抗中间相遇攻击,并具有较高的安全冗余。