With the rapid advancement of neural language models, the deployment of overparameterized models has surged, increasing the need for interpretable explanations comprehensible to human inspectors. Existing post-hoc interpretability methods, which often focus on unigram features of single input textual instances, fail to capture the models’ decision-making process fully. Additionally, many methods do not differentiate between decisions based on spurious correlations and those based on a holistic understanding of the input. Our paper introduces DISCO, a novel method for discovering global, rule-based explanations by identifying causal n-gram associations with model predictions. This method employs a scalable sequence mining technique to extract relevant text spans from training data, associate them with model predictions, and conduct causality checks to distill robust rules that elucidate model behavior. These rules expose potential overfitting and provide insights into misleading feature combinations. We validate DISCO through extensive testing, demonstrating its superiority over existing methods in offering comprehensive insights into complex model behaviors. Our approach successfully identifies all shortcuts manually introduced into the training data (100% detection rate on the MultiRC dataset), resulting in an 18.8% regression in model performance—a capability unmatched by any other method. Furthermore, DISCO supports interactive explanations, enabling human inspectors to distinguish spurious causes in the rule-based output. This alleviates the burden of abundant instance-wise explanations and helps assess the model’s risk when encountering out-of-distribution (OOD) data.

@en

Traffic monitoring services collect traffic reports and respond to users' traffic queries. However, the reports and queries may reveal the user's identity and location. Although different anonymization techniques have been applied to protect user privacy, a new security threat arises, namely, n-by-1 jamming attack, in which an anonymous contributing driver impersonates n drivers and uploads n normal reports by using n reporting devices. Such an attack will mislead the traffic monitoring service provider and further degrade the service quality. Existing traffic monitoring services do not support customized queries, and private information retrieval techniques cannot be applied directly in traffic monitoring. We formally define the new attack and propose a traffic monitoring scheme TraJ to defend the attack and achieve user-defined location privacy. Specifically, we bridge anonymous contributing drivers without disclosing their speed set by using private set intersection. Each RSU collects time traffic reports and structures a weighted proximity graph to filter out malicious colluding drivers. We design a user-defined privacy-preserving query method by encoding complex road network. We leverage the uploading phase from private aggregation to collect traffic conditions and allow requesting drivers to dynamically and privately query traffic conditions. We provide a formal analysis of TraJ to prove its privacy and security properties. We also construct a prototype based on a real-world dataset and Android smartphones to demonstrate its feasibility and efficiency. A formal analysis demonstrates the privacy and security properties. Extensive experiments illustrate the performance and defense efficacy.

@en

We introduce SparCAssist, a general-purpose risk assessment tool for the machine learning models trained for language tasks. It evaluates models' risk by inspecting their behavior on counterfactuals, namely out-of-distribution instances generated based on the given data instance. The counterfactuals are generated by replacing tokens in rational subsequences identified by ExPred, while the replacements are retrieved using HotFlip or the Masked-Language-Model-based algorithms. The main purpose of our system is to help the human annotators to assess the model's risk on deployment. The counterfactual instances generated during the assessment are the by-product and can be used to train more robust NLP models in the future.

@en

E-commerce platforms incorporate reputation systems that allow customers to rate suppliers following financial transactions. Existing reputation systems cannot defend the centralized server against arbitrarily tampering with the supplier’s reputation. Furthermore, they do not offer reputation access across platforms. Rates are faced with privacy leakages because rating activities are correlated with privacy (e.g., identity and rating). Meanwhile, raters could be malicious and initiate multiple rating attacks and abnormal rating attacks. Determining how to address these issues have both research and practical value. In this paper, we propose a blockchain-based privacy-preserving reputation system for e-commerce platforms named RepChain; our system allows cross-platform reputation access and anonymous and private ratings. Using RepChain, all e-commerce platforms collaborate and share users’ reputations by co-constructing a consortium blockchain and modeling the rating process as a finite state machine. In particular, we facilitate one-show anonymous credentials constructed from two-move blind signatures to protect customers’ identities and resist multiple rating attacks, leverage zero-knowledge range proof to verify the correctness of ratings and defend against abnormal rating attacks, design a secure sum computation protocol among nodes to update reputations, and verify ratings via batch processing and consensus hashes. Finally, we demonstrate the security and privacy of RepChain via a formal analysis and evaluate its performance based on Ethereum test network.@en

Vehicular networks allow billions of vehicular users to be connected to report and exchange real-time data for offering various services, such as navigation, ride-hailing, smart parking, traffic monitoring, and vehicular digital forensics. Fifth generation (5G) is a new radio access technology with greater coverage, accessibility, and higher network density. 5G-supported Vehicular Networks (5GVNs) have attracted plenty of attention from both academia and industry. Geared with new features, they are expected to revolutionize the mobility ecosystem to empower a portfolio of new services. Meanwhile, the development of such communication capabilities, along with the development of sensory devices and the enhancement of local computing powers, have lead to an inevitable reality of massive data (e.g., identity, location, and trajectory) collection from vehicular users. Unfortunately, 5GVN are still confronted with a variety of privacy threats. Such threats are targeted at users’ data, identity, location, and trajectory. If not properly handled, such threats will cause unimaginable consequences to users. In this survey, we first review the state-of-the-art of survey papers. Next, we introduce the architecture, features, and services of 5GVN, followed by the privacy objectives of 5GVN and privacy threats to 5GVN. Further, we present existing privacy-preserving solutions and analyze them in-depth. Finally, we define some future research directions to draw more attention and down-to-earth efforts into this new architecture and its privacy issues.@en