Abstract:Recommendation systems have become a key technology in mitigating information overload in the era of big data, with widespread applications in E-commerce and other fields. However, traditional centralized data collection methods expose significant risks of user privacy leakage. Federated learning enables collaborative model training across multiple data holders without the need to share raw user data, thus protecting privacy. Federated recommendation systems have gained considerable attention from both academia and industry. Existing federated recommendation algorithms place the model training process in a distributed environment, effectively avoiding the centralized storage of sensitive user data on a single server. However, these approaches still face challenges related to privacy leakage and high communication costs. To address these issues, this study proposes a communication-efficient federated recommendation algorithm based on differential privacy. The algorithm introduces a general sub-model selection strategy that strengthens privacy protection of user interaction data on the client side through a randomized response mechanism. On the server side, it employs maximum likelihood estimation to infer the true interaction frequencies of items and optimize the sub-model selection process. This strategy achieves an effective balance between privacy protection and model utility. The proposed algorithm is applicable not only to matrix factorization-based recommendation models but also to deep learning-based models, demonstrating high flexibility and adaptability across various recommendation scenarios. Furthermore, to reduce communication overhead, a global model partitioning strategy is proposed to address the complex structures and large parameter sizes of deep learning models. Differentiated optimization strategies are applied to shallow and deep networks to effectively mitigate communication costs. Theoretical analysis shows that the method satisfies differential privacy, while experimental results on real-world datasets demonstrate that the proposed approach preserves user data privacy without significantly compromising model utility, while substantially improving communication efficiency in federated recommendation systems.

Abstract:In recent years, deep learning has developed rapidly and achieved significant success in computer vision, with model evaluation and improvement remaining central concerns for researchers. However, the commonly used model comparison paradigm relies on training (or validation) and testing on closed datasets, and then identifies hard samples based on discrepancies between predictions and ground-truth labels, which provide feedback on model weaknesses and directions for improvement. This paradigm suffers from two major limitations: 1) the limited size and coverage of datasets often fail to faithfully reflect the true weaknesses of models; 2) procedures such as pretraining may introduce data leakage, resulting in potential biases in the demonstrated performance. To address these issues, this study proposes a general visual hard sample mining algorithm based on maximum discrepancy competition, which automatically mines real hard samples to reveal models’ deficiencies. The proposed algorithm follows the principle of “comparing models through competition” and optimizes the discovery of potential hard samples by jointly exploiting the intra-task and cross-task prediction dissimilarities, aiming to provide new test benchmarks for the field of computer vision in a controllable and efficient manner. Experimental results demonstrate that the constructed benchmark named GHS-CV exposes models’ weaknesses more effectively than single-task hard sample benchmarks (i.e., the semantic segmentation hard sample set SS-C and the salient object detection hard sample set SOD-C). Specifically, compared to DeepLabv3+ on SS-C, the mIoU drops by about 20% on GHS-CV, while compared to VST on SOD-C, the F_β decreases by about 36%.

Abstract:Fault localization is one of the most expensive, tedious, and time-consuming activities in software debugging, and it is also an indispensable step in software maintenance. Due to the variability of faults, fault localization is even more challenging in software product lines. Although significant progress has been made in fault localization for single-system software, research on fault localization for variability in software product lines is still insufficient. Meanwhile, existing methods face challenges such as low efficiency and poor root cause localization due to the issues of repeated generation and checking of feature interactions, as well as the propagation of faults between program statements. To address this, this study proposes an efficient and accurate fault localization method for software product lines, which performs localization at both the feature level and the statement level. At the feature level, based on observations of inclusion relationships and identical subsets between suspicious feature selection sets, the method identifies suspicious feature interactions more efficiently. At the statement level, a reduced causal model with mediator variables is used, combining causal effects and spectrum-based effects to achieve more precise fault localization. Four advanced fault localization methods for software product lines are selected, and experiments are conducted on six real-world software product line systems for comparison. The results demonstrate that the proposed method significantly outperforms other mainstream methods in terms of localization efficiency and accuracy.

Abstract:Spatio-temporal logical analysis refers to accurately expressing spatio-temporal relationships between entities using logical symbols. Traditional spatio-temporal logical analysis adopts two paradigms: closed-domain and open-domain approaches. Closed-domain methods predefine symbolic systems for representing spatio-temporal logic and then translate natural language into logical expressions. While ensuring accurate representation of spatio-temporal relationships, such methods face limitations in handling complex relationships due to the constraints of artificial definitions. Open-domain approaches extract keywords to represent spatio-temporal relationships using natural language itself. Although capable of covering complex relationships, these methods suffer from the semantic ambiguity inherent in natural language, resulting in imprecise logical representations. The purpose of this study is to convert natural language expressions of spatio-temporal relationships into logical language, enabling more precise representation of spatio-temporal information. To address the forementioned issues, this study considered the linguistic observation that spatio-temporal relationships in language are primarily expressed through localizers. By defining the semantics of localizers through logical symbols, the proposed framework aims to overcome both the insufficiency of coverage and the lack of precision. Accordingly, a spatio-temporal logical framework for localizers is established, including 1) the design of annotation specifications that define the logical expression scope of localizers and provide detailed annotation guidelines; 2) manual annotation of 6190 samples from the People’s Daily and CTB datasets to construct a task-specific corpus based on the proposed specifications; 3) application of large language models to perform logical reasoning on localizer-triggered spatio-temporal expressions, achieving an accuracy exceeding 70% based on corpus-driven inference.

Abstract:As autonomous driving applications are rapidly popularized, their safety has become the common focus of both academia and industry. Autonomous driving system (ADS) testing is an effective means for solving this problem. Currently, the mainstream testing method is the scenario-based simulation test, which evaluates the decision of ADS to be measured by simulating various elements of driving scenarios, such as roads and pedestrians. However, existing methods mainly focus on the construction and dynamic generation of critical driving scenarios, neglecting the influence of configuration changes of the vehicle itself, such as its weight and torque, on the decision-making of ADS deployed on the vehicle. To address this issue, based on the previous work SAFEVAR, this study proposes SAFEVCS, an efficient search method for safety-critical vehicle configurations. SAFEVCS employs a search algorithm to explore the vehicle configuration setting (VCS) that exposes safety vulnerabilities of ADS. Furthermore, to improve the diversity of the search results, SAFEVCS introduces fuzzing to optimize the conditions and constraints of crossover and mutation operators in search algorithms. To improve search efficiency, SAFEVCS further combines the vehicle dynamics knowledge, which achieves the self-adaption of search termination strategy and deduplication strategy. To evaluate the effectiveness and execution efficiency of SAFEVCS, the study takes SAFEVAR as the baseline for comparison and carries out extensive experiments under three driving scenarios. The experimental results show that VCS generated by SAFEVCS can effectively expose the safety vulnerabilities of ADS. In the two weather conditions of sunny and rainy days, under the simulation scenarios of pedestrians crossing the road, the obtained solution set significantly decreased the safety performance of the ADS under test, and under the same experiment environment, the simulation efficiency is increased by approximately 2.5 times.

Abstract:As the research on audio adversarial attacks advances, improving the transferability of adversarial audio across different models and ensuring its imperceptibility (that is, highly similar to the original audio in auditory perception) at the same time have become a research hotspot. This study proposes a new method called speak information attack (SIAttack) that can simultaneously improve the imperceptibility and transferability of adversarial audio. Specifically, the core idea of this method is to decouple speaker information from content information in the audio, and then apply small perturbations only to the speaker information, thereby achieving efficient attacks on the speaker recognition system under the premise of keeping the content information unchanged. The experiments on four speaker recognition models and three mainstream commercial APIs show that the audio generated by SIAttack is almost indistinguishable from the original audio, and can mislead all test models with a high success rate. Additionally, the transfer success rate on speaker recognition models can reach up to 100%.

Abstract:As a kind of graph structure with timestamps when nodes interact with each other, temporal graphs have more modeling advantages than static graphs. For example, they can detect money laundering, order brushing, equity relationships, financial fraud, and circular guarantees within a certain time interval. The cycle is the modeling of the behavior that forms a cycle in a temporal graph. Existing temporal cycle detection or mining methods mostly focus on the detection of non-decreasing complete cycles in time, but overlook the analysis and discovery of approximate cycles within a certain time interval. The discovery of such approximate cycles can detect fraudulent behavior with stronger cheating techniques. To address the problem of discovering approximate cycles that have already appeared within a certain time interval but are not fully displayed in a single source of data, this study first proposes an approximate cycle detection method based on the depth-first search, which is referred to as the baseline method (Baseline). It first mines complete cycles composed of edges satisfying non-decreasing order in each window, and then employs nodes that meet certain criteria as the start and end points of approximate cycles. In the subsequent windows, paths composed of edges within a certain time interval are mined, namely time-interval approximate cycles. To address the problems of Baseline, this study subsequently proposes an improved method for approximate cycle detection, referred to as the improved method (Improved). It first utilizes the node activity to enhance the possibility of start and end points, then improves the index features by adopting active paths and hotspots, and finally accelerates detection by employing the bidirectional search and connection from start and end points to hotspots. Extensive experiments on real and synthetic data demonstrate the efficiency and effectiveness of the proposed method.

Abstract:In recent years, recommender systems based on graph neural network (GNN) have made good use of the interaction structure of interaction data to learn user and item representations. However, existing recommendation models based on GNN often ignore the temporal information of interactions during aggregation, which makes it difficult to model the change characteristics of users’ interests. As a result, this causes overfitting of the recommendation model to data, and a lack of diversity in the recommendation results, thereby making it difficult to satisfy the more diversified needs of users. To this end, a temporal information-enhanced diversified recommendation model is proposed. First, an attention mechanism is employed to capture and fuse temporal information and interaction information from historical user-item interactions. Meanwhile, a feature disentanglement module is designed to disentangle smoothed global features from salient, highly discriminative key signals to reduce feature redundancy and improve representational clarity. Subsequently, neighbour selection is adopted to highlight inter-node differences and conduct graph convolution, with a layer attention mechanism employed to alleviate over smoothing. Finally, the learning of items in the long-tail category is enhanced by reweighting loss to improve the diversity.

Abstract:As multimodal multiobjective optimization faces challenges of reasonably defining the individual crowdedness and dynamically balancing the decision space and objective space in individual diversity calculation, there is still significant room for performance improvement in existing multimodal multiobjective optimization algorithms. To this end, this study proposes a multimodal multiobjective differential evolution algorithm based on adaptive individual diversity (MMODE-AID). First, based on the average Euclidean distance of individuals’ nearest neighbors in the decision space or objective space, the crowdedness of individuals can be defined by multiplying the relative distances between individuals, which can more reasonably measure the true crowdedness of each individual in the corresponding space. Second, based on the overall crowdedness of the decision space and objective space, the relative crowdedness of individuals in the corresponding space is obtained, which can reasonably and dynamically balance the influence of the current state of the decision space and objective space on individual diversity calculation during the evolution process, and is conducive to the sufficient search of each equivalent Pareto optimal solution set. By employing differential evolution as the basic optimization framework, MMODE-AID evaluates individual fitness based on adaptive individual diversity. Meanwhile, it can obtain a population with excellent performance in decision space distribution, objective space distribution and convergence during offspring generation and environmental selection. MMODE-AID is compared with seven advanced multimodal multiobjective optimization algorithms on 39 benchmark test problems and one real-world application problem to validate the algorithm’s performance. The experimental results demonstrate that MMODE-AID exhibits significant competitive advantages in solving multimodal multiobjective optimization problems. The source code and original experimental data of MMODE-AID are publicly available on GitHub: https://github.com/CIA-SZU/ZQ.

Abstract:As a widely employed interpreted language, Python faces performance challenges in execution efficiency. Just-in-time (JIT) compilers have been introduced to the Python ecosystem to dynamically compile bytecode into machine code, significantly improving program operation speed. However, the complex optimization strategies of JIT compilers may introduce program defects, thereby affecting program stability and reliability. Existing fuzz testing methods for Python interpreters struggle to effectively detect deep optimization defects and non-crashing defects in JIT compilers. To this end, this study proposes PjitFuzz, a coverage-guided defect detection method for Python JIT compilers. First, PjitFuzz proposes five mutation rules based on JIT optimization strategies to generate program variants that trigger the optimization strategies of Python JIT compilers. Second, a coverage-guided dynamic mutation rule selection method is designed to integrate the advantages of different mutation rules and generate diverse program variants. Third, a checksum-based code block insertion strategy is developed to effectively record changes in variable values during program execution and detect inconsistency in the output. Finally, differential testing is performed by combining different JIT compilation options to effectively detect defects in Python JIT compilers. This study compares PjitFuzz with two state-of-the-art Python interpreter fuzzing methods, FcFuzzer and IFuzzer. The experimental results show that PjitFuzz improves defect detection capability by 150% and 66.7% respectively, and outperforms existing methods in terms of code coverage by 28.23% and 15.68% respectively. For the validity rate of generated test programs, PjitFuzz outperforms the comparative methods by 42.42% and 62.74% respectively. In an eight-month experiment, PjitFuzz has discovered and reported 16 defects, 12 of which have been confirmed by developers.