Xinlei He

Heyi Zhang, Yule Liu, Xinlei He et al.

Federated learning (FL) enables collaborative model training while preserving data privacy, but its decentralized nature exposes it to client-side data poisoning attacks (DPAs) and model poisoning attacks (MPAs) that degrade global model performance. While numerous proposed defenses claim substantial effectiveness, their evaluation is typically done in isolation with limited attack strategies, raising concerns about their validity. Additionally, existing studies overlook the mutual effectiveness of defenses against both DPAs and MPAs, causing fragmentation in this field. This paper aims to provide a unified benchmark and analysis of defenses against DPAs and MPAs, clarifying the distinction between these two similar but slightly distinct domains. We present a systematic taxonomy of poisoning attacks and defense strategies, outlining their design, strengths, and limitations. Then, a unified comparative evaluation across FL algorithms and data heterogeneity is conducted to validate their individual and mutual effectiveness and derive key insights for design principles and future research. Along with the analysis, we frame our work to a unified benchmark, FLPoison, with high modularity and scalability to evaluate 15 representative poisoning attacks and 17 defense strategies, facilitating future research in this domain. Code is available at https://github.com/vio1etus/FLPoison.

Yule Liu, Zhiyuan Zhong, Yifan Liao et al.

The rising popularity of large language models (LLMs) has raised concerns about machine-generated text (MGT), particularly in academic settings, where issues like plagiarism and misinformation are prevalent. As a result, developing a highly generalizable and adaptable MGT detection system has become an urgent priority. Given that LLMs are most commonly misused in academic writing, this work investigates the generalization and adaptation capabilities of MGT detectors in three key aspects specific to academic writing: First, we construct MGT-Acedemic, a large-scale dataset comprising over 336M tokens and 749K samples. MGT-Acedemic focuses on academic writing, featuring human-written texts (HWTs) and MGTs across STEM, Humanities, and Social Sciences, paired with an extensible code framework for efficient benchmarking. Second, we benchmark the performance of various detectors for binary classification and attribution tasks in both in-domain and cross-domain settings. This benchmark reveals the often-overlooked challenges of attribution tasks. Third, we introduce a novel attribution task where models have to adapt to new classes over time without (or with very limited) access to prior training data in both few-shot and many-shot scenarios. We implement eight different adapting techniques to improve the performance and highlight the inherent complexity of the task. Our findings provide insights into the generalization and adaptation ability of MGT detectors across diverse scenarios and lay the foundation for building robust, adaptive detection systems. The code framework is available at https://github.com/Y-L-LIU/MGTBench-2.0.

Yule Liu, Zhen Sun, Xinlei He et al.

Recent advancements in fine-tuning proprietary language models enable customized applications across various domains but also introduce two major challenges: high resource demands and security risks. Regarding resource demands, recent work proposes novel partial compression, such as BitDelta, to quantize the delta weights between the fine-tuned model and base model. Regarding the security risks, user-defined fine-tuning can introduce security vulnerabilities, such as alignment issues, backdoor attacks, and hallucinations. However, most of the current efforts in security assessment focus on the full-precision or full-compression models, it is not well-discussed how the partial compression methods affect security concerns. To bridge this gap, we evaluate the robustness of delta-weight quantization against these security threats. In this paper, we uncover a "free lunch" phenomenon: partial compression can enhance model security against fine-tuning-based attacks with bearable utility loss. Using Llama-2-7b-chat as a case study, we show that, with under 10% utility degradation, the partial compression mitigates alignment-breaking risks by up to 66.17%, harmful backdoor vulnerabilities by 64.46%, and targeted output manipulation risks by up to 90.53%. We further apply LogitLens to visualize internal state transformations during forward passes, suggesting mechanisms for both security failure and recovery in standard versus compressed fine-tuning. This work offers new insights into selecting effective delta compression methods for secure, resource-efficient multi-tenant services.

Zhiyuan Zhong, Zhen Sun, Yepang Liu et al.

Vision Language Models (VLMs) have shown remarkable performance, but are also vulnerable to backdoor attacks whereby the adversary can manipulate the model's outputs through hidden triggers. Prior attacks primarily rely on single-modality triggers, leaving the crucial cross-modal fusion nature of VLMs largely unexplored. Unlike prior work, we identify a novel attack surface that leverages cross-modal semantic mismatches as implicit triggers. Based on this insight, we propose BadSem (Backdoor Attack with Semantic Manipulation), a data poisoning attack that injects stealthy backdoors by deliberately misaligning image-text pairs during training. To perform the attack, we construct SIMBad, a dataset tailored for semantic manipulation involving color and object attributes. Extensive experiments across four widely used VLMs show that BadSem achieves over 98% average ASR, generalizes well to out-of-distribution datasets, and can transfer across poisoning modalities. Our detailed analysis using attention visualization shows that backdoored models focus on semantically sensitive regions under mismatched conditions while maintaining normal behavior on clean inputs. To mitigate the attack, we try two defense strategies based on system prompt and supervised fine-tuning but find that both of them fail to mitigate the semantic backdoor. Our findings highlight the urgent need to address semantic vulnerabilities in VLMs for their safer deployment.

Wenhan Dong, Tianyi Hu, Jingyi Zheng et al.

Multi-round incomplete information tasks are crucial for evaluating the lateral thinking capabilities of large language models (LLMs). Currently, research primarily relies on multiple benchmarks and automated evaluation metrics to assess these abilities. However, our study reveals novel insights into the limitations of existing methods, as they often yield misleading results that fail to uncover key issues, such as shortcut-taking behaviors, rigid patterns, and premature task termination. These issues obscure the true reasoning capabilities of LLMs and undermine the reliability of evaluations. To address these limitations, we propose a refined set of evaluation standards, including inspection of reasoning paths, diversified assessment metrics, and comparative analyses with human performance.

Jingyi Zheng, Junfeng Wang, Zhen Sun et al.

As Large Language Models (LLMs) advance, Machine-Generated Texts (MGTs) have become increasingly fluent, high-quality, and informative. Existing wide-range MGT detectors are designed to identify MGTs to prevent the spread of plagiarism and misinformation. However, adversaries attempt to humanize MGTs to evade detection (named evading attacks), which requires only minor modifications to bypass MGT detectors. Unfortunately, existing attacks generally lack a unified and comprehensive evaluation framework, as they are assessed using different experimental settings, model architectures, and datasets. To fill this gap, we introduce the Text-Humanization Benchmark (TH-Bench), the first comprehensive benchmark to evaluate evading attacks against MGT detectors. TH-Bench evaluates attacks across three key dimensions: evading effectiveness, text quality, and computational overhead. Our extensive experiments evaluate 6 state-of-the-art attacks against 13 MGT detectors across 6 datasets, spanning 19 domains and generated by 11 widely used LLMs. Our findings reveal that no single evading attack excels across all three dimensions. Through in-depth analysis, we highlight the strengths and limitations of different attacks. More importantly, we identify a trade-off among three dimensions and propose two optimization insights. Through preliminary experiments, we validate their correctness and effectiveness, offering potential directions for future research.

Yifan Liao, Zhen Sun, Xiaoyun Qiu et al.

Visual Language Models (VLMs), with powerful multimodal reasoning capabilities, are gradually integrated into autonomous driving by several automobile manufacturers to enhance planning capability in challenging environments. However, the trajectory planning capability of VLMs in work zones, which often include irregular layouts, temporary traffic control, and dynamically changing geometric structures, is still unexplored. To bridge this gap, we conduct the \textit{first} systematic study of VLMs for work zone trajectory planning, revealing that mainstream VLMs fail to generate correct trajectories in $68.0%$ of cases. To better understand these failures, we first identify candidate patterns via subgraph mining and clustering analysis, and then confirm the validity of $8$ common failure patterns through human verification. Building on these findings, we propose REACT-Drive, a trajectory planning framework that integrates VLMs with Retrieval-Augmented Generation (RAG). Specifically, REACT-Drive leverages VLMs to convert prior failure cases into constraint rules and executable trajectory planning code, while RAG retrieves similar patterns in new scenarios to guide trajectory generation. Experimental results on the ROADWork dataset show that REACT-Drive yields a reduction of around $3\times$ in average displacement error relative to VLM baselines under evaluation with Qwen2.5-VL. In addition, REACT-Drive yields the lowest inference time ($0.58$s) compared with other methods such as fine-tuning ($17.90$s). We further conduct experiments using a real vehicle in 15 work zone scenarios in the physical world, demonstrating the strong practicality of REACT-Drive.

Delong Ran, Xinlei He, Tianshuo Cong et al.

Language Models (LMs) typically adhere to a "pre-training and fine-tuning" paradigm, where a universal pre-trained model can be fine-tuned to cater to various specialized domains. Low-Rank Adaptation (LoRA) has gained the most widespread use in LM fine-tuning due to its lightweight computational cost and remarkable performance. Because the proportion of parameters tuned by LoRA is relatively small, there might be a misleading impression that the LoRA fine-tuning data is invulnerable to Membership Inference Attacks (MIAs). However, we identify that utilizing the pre-trained model can induce more information leakage, which is neglected by existing MIAs. Therefore, we introduce LoRA-Leak, a holistic evaluation framework for MIAs against the fine-tuning datasets of LMs. LoRA-Leak incorporates fifteen membership inference attacks, including ten existing MIAs, and five improved MIAs that leverage the pre-trained model as a reference. In experiments, we apply LoRA-Leak to three advanced LMs across three popular natural language processing tasks, demonstrating that LoRA-based fine-tuned LMs are still vulnerable to MIAs (e.g., 0.775 AUC under conservative fine-tuning settings). We also applied LoRA-Leak to different fine-tuning settings to understand the resulting privacy risks. We further explore four defenses and find that only dropout and excluding specific LM layers during fine-tuning effectively mitigate MIA risks while maintaining utility. We highlight that under the "pre-training and fine-tuning" paradigm, the existence of the pre-trained model makes MIA a more severe risk for LoRA-based LMs. We hope that our findings can provide guidance on data privacy protection for specialized LM providers.

Wenhan Dong, Chao Lin, Xinlei He et al.

Privacy-preserving federated learning (PPFL) aims to train a global model for multiple clients while maintaining their data privacy. However, current PPFL protocols exhibit one or more of the following insufficiencies: considerable degradation in accuracy, the requirement for sharing keys, and cooperation during the key generation or decryption processes. As a mitigation, we develop the first protocol that utilizes neural networks to implement PPFL, as well as incorporating an Aggregatable Hybrid Encryption scheme tailored to the needs of PPFL. We name these networks as Homomorphic Adversarial Networks (HANs) which demonstrate that neural networks are capable of performing tasks similar to multi-key homomorphic encryption (MK-HE) while solving the problems of key distribution and collaborative decryption. Our experiments show that HANs are robust against privacy attacks. Compared with non-private federated learning, experiments conducted on multiple datasets demonstrate that HANs exhibit a negligible accuracy loss (at most 1.35%). Compared to traditional MK-HE schemes, HANs increase encryption aggregation speed by 6,075 times while incurring a 29.2 times increase in communication overhead.