Number Of Rounds Research Articles

FedLGA: Toward System-Heterogeneity of Federated Learning via Local Gradient Approximation.

Federated learning (FL) is a decentralized machine learning architecture, which leverages a large number of remote devices to learn a joint model with distributed training data. However, the system-heterogeneity is one major challenge in an FL network to achieve robust distributed learning performance, which comes from two aspects: 1) device-heterogeneity due to the diverse computational capacity among devices and 2) data-heterogeneity due to the nonidentically distributed data across the network. Prior studies addressing the heterogeneous FL issue, for example, FedProx, lack formalization and it remains an open problem. This work first formalizes the system-heterogeneous FL problem and proposes a new algorithm, called federated local gradient approximation (FedLGA), to address this problem by bridging the divergence of local model updates via gradient approximation. To achieve this, FedLGA provides an alternated Hessian estimation method, which only requires extra linear complexity on the aggregator. Theoretically, we show that with a device-heterogeneous ratio ρ , FedLGA achieves convergence rates on non-i.i.d. distributed FL training data for the nonconvex optimization problems with O ([(1+ρ)/√{ENT}] + 1/T) and O ([(1+ρ)√E/√{TK}] + 1/T) for full and partial device participation, respectively, where E is the number of local learning epoch, T is the number of total communication round, N is the total device number, and K is the number of the selected device in one communication round under partially participation scheme. The results of comprehensive experiments on multiple datasets indicate that FedLGA can effectively address the system-heterogeneous problem and outperform current FL methods. Specifically, the performance against the CIFAR-10 dataset shows that, compared with FedAvg, FedLGA improves the model's best testing accuracy from 60.91% to 64.44%.

Federated learning on low-power Arduino Nano33 BLE Sense to predict the length of stay using a linear regression model

Federated learning (FL) is a collaborative learning paradigm where multiple clients are used to build the model without sharing data and preserving privacy. An FL-based linear regression model is designed to predict the length of stay for patients at hospitals using the low-power Arduino Nano 33 BLE Sense microcontroller unit (MCU). FL uses a distributed learning technique that allows model building from decentralized data sources. The Arduino Nano 33 BLE Sense is a compact and energy-efficient MCU providing an ideal platform for implementing FL in resource-constrained environments. FL algorithms aggregate model parameters from multiple Arduino clients and collectively train and build a predictive model to estimate the length of stay at the hospital by patients. Experiments were conducted to understand the performance of FL on clients with data of equal and varying sizes and heterogeneous data from multiple sources. The performance of the algorithm is evaluated based on Mean Absolute Error (MAE), Percentage Decrease in Training error (PDTE), and Percentage Difference with Optimal Testing (PDOT) value. Experimental results show that the number of local epochs and FL rounds affects the convergence of clients to the optimal value. The experimental results demonstrate the applicability of FL on low-power MCUs, preserving privacy which is a core requirement for healthcare solutions.

Quantifying Human Naïve B Cell Proliferation Kinetics and Differentiation in Controlled In Vitro Cell Culture.

Division tracking dyes like Cell Trace Violet (CTV) enable the quantification of cell proliferation, division, and survival kinetics of human naïve B cell responses in vitro. Human naïve B cells exhibit distinct responses to different stimuli, with CpG and anti-Ig inducing a T cell-independent (TI) response, while CD40L and IL-21 promote a T cell-dependent (TD) response that induces isotype switching and differentiation into antibody-secreting cells (ASCs). Both stimulation methods yield valuable insights into the intrinsic programming of B cell health within individuals, making them useful for clinical investigations. For instance, quantitative analysis from these B cell populations could reveal biologically meaningful measurements such as the average number of division rounds and the time to cells' fate. Here, we describe a novel in vitro culture setup for CTV-labelled human naïve B cells and a method for obtaining precise time-based data on proliferation, division-linked isotype switching, and differentiation.

A Survey on Mental Health Status in West of Iran: A Spatial Analysis

"Mental disorders clearly are among the major challenges of global health causing many concerns for health systems, health experts, and health policy makers. The present study was conducted to evaluate the mental health program in comprehensive health centers and health houses under the coverage of the Ilam University of Medical Sciences in the West of Iran. The present study was conducted based on the reports obtained from the mental health program in 2019. The data were recorded into registration and evaluation forms of mental disorders with the codes 102-19-03 and 102-19-01 and then entered into the Excel software. The prevalence per 100 and 100,000 people, as well as their 95% confidence intervals, were calculated using Formula 1 ROUND (Number of cases*100000/population) and Formula 2 ROUND (Prevalence ±1.96*SQRT (Prevalence *(1- Prevalence)/population). The results showed that the prevalence of depression was 19,100 [95% CI: 18999-19201] 19100), and the prevalence of anxiety was 304,66 [95% CI: 20348-30584] per 100,000 population. The highest prevalence among mental disorders was related to depression and anxiety. Considering the high average score of mental health status in people over 65 years old, it is necessary to implement measures and interventions to improve these individuals’ mental conditions and upgrade their mental health."

Exploration and Research on Chinese Culture in University English Translation Teaching in the Background of Internet+

Abstract Under the background of Internet+, for the problems of high comprehension difficulty and inaccurate translation in university English translation teaching, this paper proposes the skeleton knowledge translation model to assist university English translation teaching that integrates Chinese culture. Starting from the objectives of university English translation teaching, Chinese culture is explored in university English translation materials to complete the design of university English translation teaching that integrates Chinese culture. Propose the application of Internet technology in the above teaching design, utilize the IBM word alignment model and TF-IDF similarity to preprocess the university English translation textbooks, and on the basis of preprocessing, estimate the probability and decode the generated phrase pairs, and then complete the construction of the English translation model based on skeleton knowledge. The English translation model’s performance analysis is carried out using skeleton knowledge after determining the data set and evaluation criteria. The results show that for the whole 400 000 batch training, the training time of the English translation model based on skeleton knowledge is about one-fifth less than that of the unsupervised neural machine translation model, indicating that with the increase of the number of rounds, the performance of the English translation model based on skeleton knowledge is also better. This study serves as a guide for exploring and researching Chinese culture in university English translation teaching and promotes the development of university English translation teaching in colleges and universities.

Resource-Aware Federated Neural Architecture Search over Heterogeneous Mobile Devices

Federated learning has been recently proposed for many clients to collaboratively train a machine learning model in a privacy-preserving manner. However, it also amplifies the difficulty in designing good neural network architecture, especially considering heterogeneous mobile devices. To this end, we propose a novel neural architecture search algorithm, namely FedNAS, which can automatically generate a set of optimal models under federated settings. The main idea is to decouple the two primary steps of the NAS process, i.e, model search and model training, and separately distribute them on the cloud and devices. It also tackles the primary challenge of limited on-device computational and communication resources through its novel designs: FedNAS fully exploits the key opportunity of insufficient model candidate re-training during the architecture search process and incorporates three key optimizations: parallel candidate training on partial clients, early dropping candidates with inferior performance, and dynamic round numbers. Evaluated on typical CNN architectures and large-scale datasets, FedNAS is able to achieve comparable model accuracy as a state-of-the-art NAS algorithm that trains models with centralized data, and also reduces the client cost by up to 200_ or more compared to a straightforward design of federated NAS.

Quantification and precision: a brief look at some ancient accounts

ABSTRACT We explore the extent to which ancient Greek authors formulated concepts that approximate or encompass our modern notions of precision and accuracy. First, we focus on estimates and measurements of geographic features, astronomical times and positions, and weight. These raise further questions about whether the quantities reported were measured, estimated, or rounded. While ancient sources discuss the use of instruments, it is not always clear that the aim was to achieve what we would today regard as ‘precision’. Next, we briefly consider round numbers, observing that they could carry symbolic meaning, while unrounded numbers could give an impression of hard-won achievement. Finally, we examine uses of the word akribeia. This is often translated as ‘precision’ or ‘exactness’, and Greek writers sometimes used akribeia to denote an ideal for their inquiries. A brief look at its uses by a number of Greek writers will on the one hand show the mismatch with our term ‘precision’, and on the other hand throw some light on the aims of Greek investigators.

HeFUN: Homomorphic Encryption for Unconstrained Secure Neural Network Inference

Homomorphic encryption (HE) has emerged as a pivotal technology for secure neural network inference (SNNI), offering privacy-preserving computations on encrypted data. Despite active developments in this field, HE-based SNNI frameworks are impeded by three inherent limitations. Firstly, they cannot evaluate non-linear functions such as ReLU, the most widely adopted activation function in neural networks. Secondly, the permitted number of homomorphic operations on ciphertexts is bounded, consequently limiting the depth of neural networks that can be evaluated. Thirdly, the computational overhead associated with HE is prohibitively high, particularly for deep neural networks. In this paper, we introduce a novel paradigm designed to address the three limitations of HE-based SNNI. Our approach is an interactive approach that is solely based on HE, called iLHE. Utilizing the idea of iLHE, we present two protocols: ReLU, which facilitates the direct evaluation of the ReLU function on encrypted data, tackling the first limitation, and HeRefresh, which extends the feasible depth of neural network computations and mitigates the computational overhead, thereby addressing the second and third limitations. Based on HeReLU and HeRefresh protocols, we build a new framework for SNNI, named HeFUN. We prove that our protocols and the HeFUN framework are secure in the semi-honest security model. Empirical evaluations demonstrate that HeFUN surpasses current HE-based SNNI frameworks in multiple aspects, including security, accuracy, the number of communication rounds, and inference latency. Specifically, for a convolutional neural network with four layers on the MNIST dataset, HeFUN achieves 99.16% accuracy with an inference latency of 1.501 s, surpassing the popular HE-based framework CryptoNets proposed by Gilad-Bachrach, which achieves 98.52% accuracy with an inference latency of 3.479 s.

PSO‐GA‐SVRmodel for S‐parameters of radio‐frequency power amplifier under different temperatures

AbstractSupport vector machine (SVR) has been introduced into the modeling of S‐parameters in radio‐frequency (RF) power amplifiers (PA). The modeling accuracy and speed of SVR are primarily affected by the penalty parametercand the kernel function coefficientγ. Using the traditional grid search technique to determine these two parameters is time‐consuming and labor‐intensive, and ensuring the model's accuracy is not easy. This article proposes an S‐parameters modeling method based on PSO‐GA‐SVR to improve the SVR's modeling accuracy and speed. The model mainly focuses on particle swarm optimization (PSO) and combines selection, crossover, and mutation operations in genetic algorithms (GA). The fitness values are arranged from small to large in each iteration process, and the first 1/3 are selected for crossover and mutation. Then, the resulting new particle swarms are introduced into the original particle swarm population for searching. On the one hand, PSO‐GA extends the population size and reduces the possibility of falling into local optimization. On the other hand, due to population size expansion, the number of iteration rounds is reduced, and the modeling speed is also increased. The experimental results show that compared to SVR, GA‐SVR, and PSO‐SVR, the proposed PSO‐GA‐SVR can improve the modeling accuracy by more than one magnitude or more while also increasing modeling speed by one magnitude or more. Furthermore, compared with the classical machine learning algorithms such as gradient boosting, random forest, and gcForset, the proposed PSO‐GA‐SVR improves the modeling accuracy by one order of magnitude and the modeling speed by two orders of magnitude more than gradient boosting, random forest, and improves the modeling speed by one order of magnitude more than gcForset.

Similarity Property and Slide Attack of Block Cipher FESH

This paper focuses on similarity properties and extension of the classical slide property of block ciphers. Taking FESH, an award-winning block cipher of the National Cryptographic Algorithm Design Competition 2019, as an example, similarity properties of the encryption and key transformation are found, owing to the similar structures that the encryption and key transformation adopted, and the constants generation. Based on the similarity properties, extended slide properties can be constructed for FESH. Slide attacks of FESH are then proposed. The similarity properties and extended slide property are immune to the increasing of iterated rounds, i.e., it cannot be avoided by increasing the round number of FESH. Furthermore, extended slide property helps relaxing the strict requirements of the subkeys in slide attacks. Taking Feistel and SPN structures as examples, frameworks of slide attacks based on the extended slide properties are presented. Slide attack of FESH is exactly a concrete example of SPN structure.

SHED: A Newton-type algorithm for federated learning based on incremental Hessian eigenvector sharing

There is a growing interest in the distributed optimization framework that goes under the name of Federated Learning (FL). In particular, much attention is being turned to FL scenarios where the network is strongly heterogeneous in terms of communication resources (e.g., bandwidth) and data distribution. In these cases, communication between local machines (agents) and the central server (Master) is a main consideration. In this work, we present SHED, an original communication-constrained Newton-type (NT) algorithm designed to accelerate FL in such scenarios. SHED is by design robust to non independent identically distributed (non i.i.d.) data distributions, handles heterogeneity of agents’ communication resources (CRs), only requires sporadic Hessian computations, and achieves global asymptotic super-linear convergence. This is possible thanks to an incremental strategy, based on eigendecomposition of the local Hessian matrices, which exploits (possibly) outdated second-order information. SHED is thoroughly validated on real datasets by assessing (i) the number of communication rounds required for convergence, (ii) the overall amount of data transmitted and (iii) the number of local Hessian computations. For all these metrics, SHED shows superior performance against state-of-the art techniques like BFGS, GIANT and FedNL.

On Large Tweaks in Tweakable Even-Mansour with Linear Tweak and Key Mixing

In this paper, we provide the first analysis of the Iterated Tweakable Even-Mansour cipher with linear tweak and key (or tweakey) mixing, henceforth referred as TEML, for an arbitrary tweak(ey) size kn for all k ≥ 1, and arbitrary number of rounds r ≥ 2. Note that TEML captures the high-level design paradigm of most of the existing tweakable block ciphers (TBCs), including SKINNY, Deoxys, TweGIFT, TweAES etc. from a provable security point of view. At ASIACRYPT 2015, Cogliati and Seurin initiated the study of TEML by showing that 4-round TEML with a 2n-bit uniform at random key, and n-bit tweak is secure up to 22n/3 queries. In this work, we extend this line of research in two directions. First, we propose a necessary and sufficient class of linear tweakey schedules to absorb mn-bit tweak(ey) material in a minimal number of rounds, for all m ≥ 1. Second, we give a rigorous provable security treatment for r-round TEML, for all r ≥ 2. In particular, we first show that the 2r-round TEML with a (2r + 1)n-bit key, αn-bit tweak, and a special class of tweakey schedule is IND-CCA secure up to O(2r−α/r n) queries. Our proof crucially relies on the use of the coupling technique to upper-bound the statistical distance of the outputs of TEML cipher from the uniform distribution. Our main echnical contribution is a novel approach for computing the probability of failure in coupling, which could be of independent interest for deriving tighter bounds in coupling-based security proofs. Next, we shift our focus to the chosen-key setting, and show that (r + 3)-round TEML, with rn bits of tweakey material and a special class of tweakey schedule, offers some form of resistance to chosen-key attacks. We prove this by showing that r + 3 rounds of TEML are both necessary and sufficient for sequential indifferentiability. As a consequence of our results, we provide a sound provable security footing for the TWEAKEY framework, a high level design rationale of popular TBC.

Improved Attacks on LowMC with Algebraic Techniques

The LowMC family of SPN block cipher proposed by Albrecht et al. was designed specifically for MPC-/FHE-/ZKP-friendly use cases. It is especially used as the underlying block cipher of PICNIC, one of the alternate third-round candidate digital signature algorithms for NIST post-quantum cryptography standardization. The security of PICNIC is highly related to the difficulty of recovering the secret key of LowMC from a given plaintext/ciphertext pair, which raises new challenges for security evaluation under extremely low data complexity.In this paper, we improve the attacks on LowMC under low data complexity, i.e. 1 or 2 chosen plaintext/ciphertext pairs. For the difference enumeration attack with 2 chosen plaintexts, we propose new algebraic methods to better exploit the nonlinear relation inside the introduced variables based on the attack framework proposed by Liu et al. at ASIACRYPT 2022. With this technique, we significantly extend the number of attack rounds for LowMC with partial nonlinear layers and improve the success probability from around 0.5 to over 0.9. The security margin of some instances can be reduced to only 3/4 rounds. For the key-recovery attack using a single plaintext, we adopt a different linearization strategy to reduce the huge memory consumption caused by the polynomial methods for solving multivariate equation systems. The memory complexity reduces drastically for all 5-/6-round LowMC instances with full nonlinear layers at the sacrifice of a small factor of time complexity. For 5-round LowMC instances with a block size of 129, the memory complexity decreases from 286.46 bits to 248.18 bits while the time complexity even slightly reduces. Our results indicate that the security for different instances of LowMC under extremely low data complexity still needs further exploration.

WAS: improved white-box cryptographic algorithm over AS iteration

The attacker in white-box model has full access to software implementation of a cryptographic algorithm and full control over its execution environment. In order to solve the issues of high storage cost and inadequate security about most current white-box cryptographic schemes, WAS, an improved white-box cryptographic algorithm over AS iteration is proposed. This scheme utilizes the AS iterative structure to construct a lookup table with a five-layer ASASA structure, and the maximum distance separable matrix is used as a linear layer to achieve complete diffusion in a small number of rounds. Attackers can be prevented from recovering the key under black-box model. The length of nonlinear layer S and affine layer A in lookup table is 16 bits, which effectively avoids decomposition attack against the ASASA structure and makes the algorithm possess anti-key extraction security under the white-box model, while WAS possesses weak white-box (32 KB, 112)-space hardness to satisfy anti-code lifting security. WAS has provable security and better storage cost than existing schemes, with the same anti-key extraction security and anti-code lifting security, only 128 KB of memory space is required in WAS, which is only 14% of SPACE-16 algorithm and 33% of Yoroi-16 algorithm.

Tighter Trail Bounds for Xoodoo

Determining bounds on the differential probability of differential trails and the squared correlation contribution of linear trails forms an important part of the security evaluation of a permutation. For Xoodoo, such bounds were proven using the trail core tree search technique, with a dedicated tool (XooTools) that scans the space of all r-round trails with weight below a given threshold Tr. The search space grows exponentially with the value of Tr and XooTools appeared to have reached its limit, requiring huge amounts of CPU time to push the bounds a little further. The bottleneck was the phase called trail extension where short trails are extended to more rounds, especially in the backward direction. In this work, we present a number of techniques that allowed us to make extension much more efficient and as such to increase the bounds significantly. Notably, we prove that the minimum weight of any 4-round trail is 80, the minimum weight of any 6-round trail is at least 132 and the minimum weight of any 12-round trail is at least 264, both for differential and linear trails. As a byproduct we found families of trails that have predictable weight once extended to more rounds and use them to compute upper bounds for the minimum weight of trails for arbitrary numbers of rounds.

Committing Security of Ascon: Cryptanalysis on Primitive and Proof on Mode

Context-committing security of authenticated encryption (AE) that prevents ciphertexts from being decrypted with distinct decryption contexts, (K,N,A) comprising a key K, a nonce N, and associate data A is an active research field motivated by several real-world attacks. In this paper, we study the context-committing security of Ascon, the lightweight permutation-based AE selected by the NIST LWC in 2023, for cryptanalysis on primitive and proof on mode. The attacker’s goal is to find a collision of a ciphertext and a tag with distinct decryption contexts in which an attacker can control all the parameters including the key. First, we propose new attacks with primitives that inject differences in N and A. The new attack on Ascon-128 improves the number of rounds from 2 to 3 and practically generates distinct decryption contexts. The new attack also works in a practical complexity on 3 rounds of Ascon-128a. Second, we prove the context-committing security of Ascon with zero padding, namely Ascon-zp, in the random permutation model. Ascon-zp achieves min {t+z/2 , n+t−k−ν/2 , c/2}-bit security with a t-bit tag, a z-bit padding, an n-bit state, a ν-bit nonce, and a c-bit inner part. This bound corresponds to min {64 + z/2 , 96} with Ascon-128 and Ascon-128a, and min {64 + z/2 , 80} with Ascon-80pq. The original Ascon (z = 0) achieves 64-bit security bounded by a generic birthday attack. By appending zeroes to the plaintext, the security can be enhanced up to 96 bits for Ascon-128 and Ascon-128a and 80 bits for Ascon-80pq.

Revisiting Yoyo Tricks on AES

At Asiacrypt 2017, Rønjom et al. presented key-independent distinguishers for different numbers of rounds of AES, ranging from 3 to 6 rounds, in their work titled “Yoyo Tricks with AES”. The reported data complexities for these distinguishers were 3, 4, 225.8, and 2122.83, respectively. In this work, we revisit those key-independent distinguishers and analyze their success probabilities.We show that the distinguishing algorithms provided for 5 and 6 rounds of AES in the paper of Rønjom et al. are ineffective with the proposed data complexities. Our thorough theoretical analysis has revealed that the success probability of these distinguishers for both 5-round and 6-round AES is approximately 0.5, with the corresponding data complexities mentioned earlier.We investigate the reasons behind this seemingly random behavior of those reported distinguishers. Based on our theoretical findings, we have revised the distinguishing algorithm for 5-round AES. Our revised algorithm demonstrates success probabilities of approximately 0.55 and 0.81 for 5-round AES, with data complexities of 229.95 and 230.65, respectively. We have also conducted experimental tests to validate our theoretical findings, which further support our findings.Additionally, we have theoretically demonstrated that improving the success probability of the distinguisher for 6-round AES from 0.50000 to 0.50004 would require a data complexity of 2129.15. This finding invalidates the reported distinguisher by Rønjom et al. for 6-round AES.

Propagation of Subspaces in Primitives with Monomial Sboxes: Applications to Rescue and Variants of the AES

Motivated by progress in the field of zero-knowledge proofs, so-called Arithmetization-Oriented (AO) symmetric primitives have started to appear in the literature, such as MiMC, Poseidon or Rescue. Due to the design constraints implied by this setting, these algorithms are defined using simple operations over large (possibly prime) fields. In particular, many rely on simple low-degree monomials for their non-linear layers, essentially using x ↦ x3 as an S-box.In this paper, we show that the structure of the material injected in each round (be it subkeys in a block cipher or round constants in a public permutation) could allow a specific pattern, whereby a well-defined affine space is mapped to another by the round function, and then to another, etc. Such chains of one-dimensional subspaces always exist over 2 rounds, and they can be extended to an arbitrary number of rounds, for any linear layer, provided that the round-constants are well chosen.As a consequence, for several ciphers like Rescue, or a variant of AES with a monomial Sbox, there exist some round-key sequences for which the cipher has an abnormally high differential uniformity, exceeding the size of the Sbox alphabet.Well-known security arguments, in particular based on the wide-trail strategy, have been reused in the AO setting by many designers. Unfortunately, our results show that such a traditional study may not be sufficient to guarantee security. To illustrate this, we present two new primitives (the tweakable block cipher Snare and the permutation-based hash function Stir) that are built using state-of-the-art security arguments, but which are actually deeply flawed. Indeed, the key schedule of Snare ensures the presence of a subspace chain that significantly simplifies an algebraic attack against it, and the round constants of Stir force the presence of a subspace chain aligned with the rate and capacity of the permutation. This in turns implies the existence of many easy-to-find solutions to the so-called CICO problem.

Full-round impossible differential attack on shadow block cipher

Lightweight block ciphers are the essential encryption algorithm for devices with limited resources. Its goal is to ensure the security of data transmission through resource-constrained devices. Impossible differential cryptanalysis is one of the most effective cryptanalysis on block ciphers, and assessing the ability of resisting this attack is a basic design criterion. Shadow is a lightweight block cipher proposed by Guo et al. (IEEE Internet Things J 8(16):13014–13023, 2021). It utilizes a combination of ARX operations and generalized Feistel structure to overcome the weakness of the traditional Feistel structure that only diffuses half in one round. In this paper, we focus on the differential property of Shadow and its security against impossible differential cryptanalysis. First, we use the SAT method to automatically search for a full-round impossible differential distinguisher of Shadow-32. Then, based on the experimental results, we prove that Shadow has a differential property with probability 1 based on the propagation of the state. Further, we can obtain an impossible differential distinguisher for an arbitrary number of rounds of Shadow. Finally, we perform a full key recovery attack on the full-round Shadow-32 and Shadow-64. Both experimentally and theoretically, our results indicate that Shadow is critically flawed, and regardless of the security strength of the internal components and the number of rounds applied, the overall cipher remains vulnerable to impossible differential cryptanalysis.

Research on Multi-Domain Intelligent Customer Service Dialog Modeling with Integrated Transfer Learning Strategies

Abstract The intelligent customer service dialog model is centered on human-machine dialog, which has good prospects for commercial applications in multiple domains. In this paper, we use the Siamese-LSTM model to do vectorization of questions in the FAQ question and answer database to get the semantic representation vector of sentences, and then use the approximate retrieval algorithm to index the question and answer database and perform approximate nearest-neighbor retrieval of the query. After completing the question query, migration learning is employed to create a mapping between input questions and human responses, enabling the model to produce sentences that are similar to human responses. Tests show that the task success rate gradually stabilizes around 0.80 at about the 100th round and fluctuates up to around 0.986 after that. For the average number of conversation rounds, migration learning improves the conversation efficiency of intelligent customer service, and the average number of conversation rounds gradually stabilizes at about 150 rounds and eventually stabilizes at about 4.2 rounds as the number of training rounds increases. The transfer learning strategy helps machine responses to be as close to human responses as possible.